Proceedings of the Annual Meeting of the Cognitive Science Society

All around us are traces of human design, from color-coded subway maps that facilitate navigation to furniture that balances form and function. The human capacity for creation has long fascinated cognitive scientists. Early studies of innovation highlighted the role of problem-solving, elucidating the roles of search and heuristics (Simon, 1996; Newell, 1972). Research on object perception and tool use enhanced our understanding of how humans interact with and manipulate their environment (Gibson, 1977; Norman, 1999). Subsequently, research in the visual and spatial domains uncovered key abstractions supporting reasoning, communication, and expression through visual forms, such as mental models, diagrams, and spatial analogies (Hegarty, 2011; Tversky, 2010; Goel, 1995).

This workshop focuses on information theory and cognition. The goal is to create a multidisciplinary space for discussing the most recent advances at the intersection of information theory and cognitive science and to explore how this emerging research area can help the field advance toward a more comprehensive and principled mathematical theory of human cognition.

One of the oldest and deepest questions across the cognitive sciences concerns the architecture of the mind (Fodor, 1983). Which cognitive capacities are supported by domain-general computations that apply to a broad set of inputs, and which are supported by domain-specific computations, that are specialized for a particular mental function? This question has motivated a broad set of research programs, ranging from the evolution and development of human knowledge (Tomasello, Melis, Tennie, Wyman, & Herrmann, 2012), to the specialization of neural functions (Kanwisher, 2010), to building machines to think and learn in the same ways that we can (Lake, Ullman, Tenenbaum, & Gershman, 2017).

Reasoning is one of the hallmarks of both natural and artificial intelligence. Understanding how reasoning operates in the human mind is crucial in cognitive science. Despite a long history of research on human reasoning in cognitive science—ranging from heuristics (Tversky & Kahneman, 1974) to mental models (Johnson-Laird, 1983), and from Bayesian modeling (Oaksford & Chater, 2007; Griffiths, Chater, & Tenenbaum, 2024) to neuroscience (Goel & Dolan, 2003)—little is known about how humans reason so flexibly in real life and how reasoning contributes to high-level cognitive functions including planning, social interaction, complex problem-solving, and open-ended mental exploration. Previous studies face challenges that hinder a deeper understanding of reasoning, including the difficulty of designing well-balanced experimental paradigms that maintain both control and ecological validity, efficient data collection and analysis beyond pure behavioral measures (e.g., Think-Aloud text data (Simon & Ericsson, 1984)), and understanding complex interactions between reasoning and other high-level cognitive functions, such as memory, theory of mind, and language.

People are general purpose problem solvers. We obtain food and shelter, manage companies, solve moral dilemmas, spend years toiling away at thorny math problems, and even adopt arbitrary problems through puzzles and games. The cognitive flexibility which allows us to represent and reason about such a wide range of problems, often referenced as a distinguishing feature of human intelligence (Tomasello, 2022), presents us with an especially ubiquitous one: deciding which problem to solve. The meta-level problem of what problem to choose exists, in part, because people have limited problem solving resources (Griffiths et al., 2020). While this challenge has been examined through various lenses across cognitive science, implicit in many of these perspectives is the notion of bounded rationality. Given our limited time and energy, how do we decide which problems are worthwhile and when we should quit to pursue something new?

Structure matters in cognitive science. Whether we are asking about memory retrieval, semantic representations, categorization, language acquisition, learning from complex information, aging, or creativity, cognitive scientists often find themselves forced to reckon with structure. Network science offers a quantitative approach for doing this by allowing us to ask questions about the relationships between various entities at scales ranging from dyads, to communities, to entire systems. In this case, the entities are the nodes in the network and the relationships are the edges between them. Exploring how this plays out in actual practice is incredibly varied, aesthetically and intellectually beautiful, and deeply rewarding, allowing us to develop and test hypotheses about cognition that are not otherwise possible. As a metric ruler measures length, allowing us to compare human height with the Burj Khalifa, network science measures structure, allowing us to compare the structure of our environments with the structure of our cognitive representations, how those representations change across the lifespan, and how different processes interacting with those structures generate behavior.

This workshop aims to guide cognitive science PhD students on starting a company within the start-up ecosystem., with an emphasis on how to use venture capital (VC). It will cover essential concepts, stages, decision points, and skills needed to improve chances for success. The workshop will also compare VC funding phases with those in academic research to make the concepts and processes more accessible. This workshop will begin by helping students calibrate their motivations and timelines for creating a startup company or joining one. After a review of the typical startup lifecycle by a venture capitalist, three startup founders will explain their career trajectories and critical decision points. They will then introduce important decisions about product, market, software, and funding strategies. Students will then participate in breakout sessions in one of these four areas. After the breakout sessions, the workshop will conclude with open discussion with the presenters

One of the most influential computational paradigms in modern cognitive science is the Bayesian modeling of social cognition. This paradigm models people's intuitions about other agents in terms of recursive probabilistic reasoning: agents are treated as approximately-rational decision-makers, who make Bayesian inferences about *other* agents' mental states from their observable behavior. Computational models designed in this tradition have been used in seminal work on theory-of-mind, language and communication, emotion understanding, and many other key areas of interest in cognitive science.

\section{How does the mind emerge from the brain? } A neural process theory of cognition is one of the central goals l of cognitive science. Connectionism was developed in pursuit of that goal leading to a range of proposed frameworks such as LISA and DORA \cite{DoumasEtAl2022}, SPAUN \cite{StewartEliasmith2012} and many others \cite{KotserubaTsotso2020}. The rise of neural AI reframes the question of how neural accounts may reach higher cognition \cite{SmolenskyEtAl2022,OReillyEtAl2022}. But what is meant by neural process theory ? Which neural principles would form the basis for such a theory? And what is it a theory of? What properties of cognition must such a theory address? Dynamic Field Theory (DFT) gives specific answers to these questions, postulating that the dynamics of neural populations are the level of neural processing that most closely reflects the laws of cognition. It emphasizes the emergence of cognition from the sensory-motor domain, so that a theory of cognition must address both acting and thinking.

Higher education and science are facing unprecedented challenges. What can we do as scientists? This symposium aims to highlight ongoing efforts to protect our research, institutions, and academic rights, provide concrete actions we can take, and collectively brainstorm a strategy going forward.

Large Language Models (LLMs) and Vision Language Models (VLMs) have become a dominant force in artificial intelligence and have already made a major impact on the cognitive sciences, but debate persists concerning the extent to which they possess emergent cognitive capacities. Investigation of these systems at the level of behavioral outputs has led to conflicting findings, and the question of how these outputs are generated (at a mechanistic or algorithmic level) remains open. Yet, the abilities they do exhibit behaviorally offer an unprecedented opportunity to answer longstanding questions about how neural networks could, even in principle, achieve abilities that are thought to require structured representations—such as syntactic, combinatorial, and variable-binding operations. In this symposium, we highlight a recent body of work that addresses this gap in understanding by investigating the internal mechanisms that support cognitive processing in LLMs and other large-scale neural networks. The symposium brings together researchers with backgrounds in both computer science and psychology, exploring ways in which mechanistic interpretability research and cognitive science can mutually inform one another.

Unlike in other animals that might reach full maturity within a few months or years, human cognitive development follows an unusually protracted timeline. In many contemporary societies, people might require several years of scaffolded learning opportunities to develop the full suite of cognitive skills and abilities they are expected to have as adults. This extended period of development reflects our species' unique capacity for cumulative cultural learning: humans have evolved specialized cognitive mechanisms that enable us to learn from, communicate with, and teach others across the lifespan (Csibra & Gergely, 2009; Tomasello, 2016).

The public sphere has shifted online, fundamentally altering the way political ideas are shared, debated, and contested. From topic-based chatrooms and social media threads to viral memes and videos, the internet has created new modes of political communication. To grasp how political thought is constructed in this digital landscape, we need to examine both the content of political messages and the structures that shape their impact. Studying this question is difficult not just because of the complexity of measuring and intervening on human political cognition in a noisy naturalistic setting but because of, for example, the proliferation of bots in online spaces and the frequency of insincere bad-faith posting and commentary. The research presented at this symposium advances our understanding of online political discourse by applying multiple methodologies, including statistical modeling and natural language processing of social media corpora, along with novel experiments on persuasion and how to counter misinformation.

Humans have the remarkable ability to construct commonsense theories about the world through perception. From inferring the physical relationships between objects to understanding social interactions and predicting future events, perception plays a foundational role in shaping how we learn and reason. But how do systems for perception and higher level cognitive theories interact? Are perception and common sense knowledge distinct, such that perception simply provides information for other cognitive systems to act on (Firestone & Scholl, 2016)? Or are they interlinked in ways that allow common-sense theories to affect what we perceive, or represent the world for other commonsense domains (Levin, Baker, & Banaji, 2016)?

Successfully navigating the social world requires individuals to flexibly adapt their behavior to different situational demands. Yet, the social world is also governed by broad behavioral rules: norms that prescribe or proscribe behaviors in certain situations. This poses a particularly interesting problem for early development: learning the foundational structure of norms may require strict adherence, but effective social functioning involves flexibility. Here, we refer to the ability to flexibly navigate social norms as social adaptation.

How do people understand what is normatively expected from them? And how does normative cognition motivate action? A distinct feature of human sociality is our capacity to sense the appropriate thing to say in a certain context or the action that should be done and to care enough to do it. Such competence with social norms stabilizes joint action between dyads and enables cooperation among thousands, but whether it is the product of domain-general learning mechanisms or dedicated cognitive ones is still unclear (Heyes, 2024). Moreover, as the many blunders made by contemporary AI systems like LLMs show, autonomous AI agents too need to acquire some level of normative competence to be reliable partners. However whether this is possible with current architectures is still debated (Browning, 2024). The aim of this symposium is to bring together researchers to discuss the biological, cognitive, computational and interactive bases of normative motivation and judgment.

We use language to communicate our thoughts. But is language merely the expression of thoughts, which are themselves produced by other, nonlinguistic parts of our minds? Or does language play a more transformative role in human cognition, allowing us to have thoughts that we otherwise could (or would) not have? Recent developments in artificial intelligence and cognitive science have reinvigorated this old question. Could language hold the key to the emergence of both artificial intelligence and important aspects of human intelligence? The four contributions in this symposium address this question by drawing on behavioral and neural evidence from people, and the remarkable recent developments in AI which appear to show that artificial neural networks trained on language come to have an astonishing range of abilities. Despite the diversity of the speakers' perspectives, the four contributions paint a coherent (if complex) picture: The abilities of large language models (LLMs) serve as an existence proof of just how much can—in principle—be learned from language. LLMs also act as a stress test of cognitive theories. The evidence of neural dissociation between linguistic and conceptual processing points to the multiple realizability of human-like cognition. Finally, there is an acknowledged need for systematic research on how the successes and failures of LLMs inform our understanding of human cognition.

How do children learn to talk to others? Mastery of language means being able to communicate with a wide array of interlocutors (Schieffelin & Ochs, 1986). Yet researchers have tended to treat parent-child interaction as the paradigmatic site of language learning, neglecting how children learn to use language with other people in their lives. As a result, we have come to not only lack basic facts about the full distribution of children's language input and use; we miss precisely those contexts that call for complex conversational skills, such as adapting to novel interlocutors and maintaining conversations without parent scaffolding.

Caregiving is essential to human survival and flourishing, yet it has been largely overlooked across scientific disciplines, including economics, philosophy, politics, and, importantly – cognitive science. Caregiving remains poorly understood, in part, because it does not fit neatly within traditional frameworks of human cognition and behavior (Gopnik, 2023). Take, for instance, theories of morality and cooperation (Kleiman-Weiner, Saxe, & Tenenbaum, 2017; Powell, 2022). Unlike the principle of universalism, which emphasizes impartiality, caregiving involves prioritizing the needs of specific individuals over others (Gilligan, 1993). Caregiving also directly challenges the utilitarian principle of "the greatest good for the greatest number", since it involves actions that benefit others at significant personal cost. Furthermore, caregiving diverges from the principle of reciprocity – a cornerstone of human cooperation – because caregivers generally do not expect anything in return for their actions (Fiske, 1992).

This symposium will present a set of four talks and a panel discussion that will together take the audience inside a scientific revolution that has been (largely quietly) unfolding in the field of attention: A set of recent computational modeling approaches that allow us to think about human attention in fundamentally new ways. In cognitive science, studies of attention stand out in at least two dimensions. First, and most bluntly, it is an outright confusing area to work in. "Attention" is a term ascribed to many sorts of mechanisms and phenomena. Case in point: there are at least three papers all published in 2024 presenting ongoing active (and, surprisingly, topically, largely non-overlapping) debates: Rosenholtz (2024), Theeuwes (2024), and Wu (2024). Second, attention stands out in the extent of the gap between the rich empirical phenomena integrated into conceptual theories, versus formal computational models, with most influential models dating back at least a decade (e.g., Bruce & Tsotsos, 2005; Bundesen et al., 2015; Reynolds & Heeger, 2009; Wolfe, Cave, & Franzel, 1994; Dosher & Lu, 2000), rather than keeping up with the advances in experimental work. This 90-minute-long gathering will show how the field of attention has been radically changing along both dimensions --- how models of attention have been carving new and productive ways of better drawing the contours of what attention is and enabling progress toward a more integrated research landscape of experiments and modeling.

Large Language Models (LLMs) are increasingly used in tasks such as psychological text analysis and decision-making in automated workflows. However, their reliability remains a concern due to potential biases inherited from their training process. In this study, we examine how different response format—binary versus continuous— may systematically influence LLMs' judgments. In a value statement judgments task and a text sentiment analysis task, we prompted LLMs to simulate human responses and tested both formats across several models, including both open-source and commercial models. Our findings revealed a consistent negative bias: LLMs were more likely to deliver "negative" judgments in binary formats compared to continuous ones. Control experiments further revealed that this pattern holds across both tasks. Our results highlight the importance of considering response format when applying LLMs to decision tasks, as small changes in task design can introduce systematic biases.

How do humans adapt how they reason to varying circumstances? Prior research has argued that reasoning comes in two types: a fast, intuitive type and a slow, deliberate type. Are these the only options, or can people adjust their reasoning continuously by trading off speed and accuracy within individual reasoning steps? We investigate this possibility in an experiment where participants were trained on relationships between local variables in a simple causal model, then asked to make predictions about all pairs of variables. Participants in one condition had a 5-second time limit. We found main effects of time pressure and locality, but only a small interaction in the direction opposite to our hypothesis. We present a process-level model of this phenomenon using early readouts from transformer language models. Our findings are consistent with people reasoning step by step, but accepting a higher error rate at each step under time pressure.

As we interact with the world, we learn how to allocate our attention effectively to achieve our goals. In this study, participants' eye movements were tracked as they engaged in a same-different task with novel stimuli. Participants then completed a goal-directed task using physical copies of the items. In the task, the goal-relevance of the attributes varied. This was followed by a second block of the same-different task. We analyzed how attention to the goal-relevant, goal-incidental, and goal-irrelevant attributes of the items shifted from the initial block of the same-different task to the second. The gaze patterns clearly distinguished between the goal-relevant and goal-irrelevant attributes. However, there was no distinction between the goal-relevant and goal-incidental attributes. We discuss the implications for understanding how goal-directed interactions shape attentional allocation and how this relates to what people learn about the items.

Soft production preferences are a key feature of incremental language production, yet they lack a well-defined unified explanatory theory. Here, we propose an information-theoretic theory of availability effects grounded in the notion of lossy-context working memory, which takes the form of a cost function that can be applied to any computational-level model of language production. We show that production policies that minimize this cost function naturally give rise to key soft preferences observed in empirical data, including frequency bias, heavy-NP shift, and agreement attraction. We then show a novel prediction made by the model regarding the entropy of arguments' thematic roles, and show that this effect holds in corpus data.

Word meanings are contrastive. When we are told that some- thing is a square, we are also told that it is not a triangle. How- ever, words may be learned in different contrasts. One person might learn about squares in contrast with circles. Does this mean that the two have different representations of "square?" To answer this question, participants learned to label novel shapes with novel labels in a category learning task. Critically, we manipulated the contrast participants received during learn- ing: an A-shape is specifically not a B or an A-shape is specifi- cally not a D. Afterwards, we tested participants' knowledge of the learned categories using explicit categorization tasks and similarity judgments. Contrast during learning mattered. Shapes from contrasted categories were categorized more ac- curately, were less confusable and rated as less similar.

Depression affects everyday decision-making, yet it remains unclear if such effects depend on the decision context or fluctuate over time. In this repeated-measures study, online participants completed a social exchange task (ultimatum game) and a non-social reversal learning task at baseline (n=236) and 1-month later (n=131). Mood symptoms were assessed using the Beck Depression Inventory and the Positive Valence System Scale. Psychiatric symptoms were stable over time, and dropout was unrelated to symptom severity. Mixed-effects regression revealed consistent behavioral effects of depressive symptoms—while controlling for anhedonia—across time points. Specifically, greater depressive severity predicted slower reaction times and reduced acceptance of unfair offers in the ultimatum game. An interaction between depressive and anhedonic symptoms on mood ratings emerged at baseline but did not replicate at follow-up. There were no consistent significant effects of depression on the non-social reversal learning task across time points. These findings highlight the longitudinal stability of depressive symptoms on social decision-making.

Large Language Models (LLMs) with hundreds of billions of parameters have exhibited human-like intelligence by learning from vast amounts of internet-scale data. However, the uninterpretability of large-scale neural networks raises concerns about the reliability of LLM. Studies have attempted to assess the psychometric properties of LLMs by borrowing concepts from human psychology to enhance their interpretability, but they fail to account for the fundamental differences between LLMs and humans. This results in high rejection rates when human scales are reused directly. Furthermore, these scales do not support the measurement of LLM psychological property variations in different languages. This paper introduces AIPsychoBench, a specialized benchmark tailored to assess the psychological properties of LLM. It uses a lightweight role-playing prompt to bypass LLM alignment, improving the average effective response rate from 70.12% to 90.40%. Meanwhile, the average biases are only 3.3% (positive) and 2.1% (negative), which are significantly lower than the biases of 9.8% and 6.9%, respectively, caused by traditional jailbreak prompts. Furthermore, among the total of 112 psychometric subcategories, the score deviations for seven languages compared to English ranged from 5% to 20.2% in 43 subcategories, providing the first comprehensive evidence of the linguistic impact on the psychometrics of LLM.

With the surge in cognitive gaming data, understanding players' learning patterns and cognitive growth has become increasingly important. These data offer valuable opportunities to study individual cognitive development during learning. However, the diversity of player profiles and the complexity of gaming tasks pose significant challenges for accurate skill prediction. Specifically, the heterogeneity of player profiles leads to diverse and complex learning trajectories; data sparsity and temporal dynamics further exacerbate these challenges. To address these challenges, we propose the MFCGT (Multi-feature Forget Cognitive Game Tracing) model. First, we perform multi-feature selection to extract key features from player behavior data to reduce noise and improve prediction accuracy. Second, we introduce a time-aware decay mechanism that simulates skill degradation using an exponential decay function, ensuring the model captures the impact of forgetting on learning trajectories. Finally, we incorporate an attention mechanism to dynamically identify the most relevant historical performance for the current task, thereby enhancing the model's predictive capability. The experiment results show that MFCGT significantly outperforms traditional models in skill prediction tasks. Additionally, MFCGT effectively captures players' learning dynamics and forgetting effects, providing more accurate learning predictions.

Speakers often communicate imprecisely, using expressions that are strictly speaking false yet felicitous. The degree to which imprecision is tolerated in discourse is governed by the Standard of Precision (SoP). While it is known that contextual factors can modulate the SoP, less is known about the discourse dynamics of imprecision. Previous accounts have claimed that implicit negotiations of the SoP are unidirectional: they only work upward, not downward. Here, we investigated whether parallel asymmetries arise when comprehenders adapt their SoPs in response to an interlocutor's precision preferences. Results show bidirectional SoP adaptation effects: exposure to lower standards increased tolerance for imprecision, while exposure to higher standards reinforced stricter thresholds. These updates persisted beyond the dialogue, suggesting that exposure to (im)precise speakers modulates not only interpretations within a discourse, but also beyond the conversation. More broadly, our study provides a novel framework for studying real-time dynamic meaning negotiations during conversation.

Understanding social scenes depends on tracking relational visual information, which is prioritized behaviorally and represented in the superior temporal sulcus (STS), a region involved in processing social scenes. Despite its importance, relational information has been underutilized in computational models of social vision. In this study, we evaluate two neural network models—SocialGNN and RNN Edge—that explicitly incorporate relational cues, and compare their performance to state-of-the-art (SOTA) AI vision models. SocialGNN utilizes a graph neural network to organize input information about each video frame into a graph structure with nodes representing faces and key objects, and edges encoding relational information such as gaze direction and physical contact. RNN Edge is an even simpler model that processes only relational information without node features or graph-based structures. These models were tested on behavioral and neural data from 3-second natural videos of two people engaged in everyday activities, as well as on the PHASE dataset, a collection of 2D animations depicting agent-object interactions inspired by Heider and Simmel. Across both datasets, SocialGNN and RNN Edge achieved strong performance in predicting human behavioral ratings of social interactions and were comparable to SOTA AI models in behavioral encoding tasks, despite being trained on significantly less data and with simpler architectures. Notably, the success of RNN Edge suggests that additional visual features and the graph-based framework of SocialGNN do not significantly enhance performance, underscoring the primacy of gaze and physical contact as essential relational cues. These findings emphasize the importance of integrating relational information into computational models to develop better models of social perception and human-aligned AI.

Generalizations are powerful tools that agents rely on to predict and control their environments. However, some generalizations are restricted to "sociocultural bubbles" (e.g. "women have trouble getting tenure in math"). How are such patterns communicated? We report one interdisciplinary study — bridging philosophy, linguistics, and cognitive developmental psychology — which examined the developing capacity for contextual restriction of generics in 4-7-year-olds and adults (N=200). We provided context cues signaling that the speaker used a generic generalization to convey a broad vs. contextually-restricted regularity, and measured endorsement of generics attributing properties prevalent globally vs. within "bubbles". Adults endorsed generics flexibly, tracking context cues, but younger children struggled, over-attributing socially-contingent properties to the group beyond the "bubble", on par with context-general regularities. This reveals a troubling discrepancy between children and adults' interpretations of generics, opening the door for miscommunication. We discuss strategies to mitigate this in educational and family communication settings.

Music evokes profound emotions, yet the universality of emotional descriptors across languages remains debated. A key challenge in cross-cultural research on music emotion is biased stimulus selection and manual curation of taxonomies, predominantly relying on Western music and languages. To address this, we propose a balanced experimental design with nine online experiments in Brazil, the US, and South Korea, involving N=672 participants. First, we sample a balanced set of popular music from these countries. Using an open-ended tagging pipeline, we then gather emotion terms to create culture-specific taxonomies. Finally, using these bottom-up taxonomies, participants rate emotions of each song. This allows us to map emotional similarities within and across cultures. Results show consistency in high arousal, high valence emotions but greater variability in others. Notably, machine translations were often inadequate to capture music-specific meanings. These findings together highlight the need for a domain-sensitive, open-ended, bottom-up emotion elicitation approach to reduce cultural biases in emotion research.

Social routine words – e.g. yes, no, hi, bye, okay, and thank you – are among the first words children learn across different languages and they help constitute a foundational set of actions for conducting social interactions. Despite this, we know little about how these words are acquired. In this paper we begin by showing that social routine words are systematically acquired earlier than statistical models of word acquisition predict. Furthermore, we argue this gap is due to a selective focus of word learning science on reference – how words are mapped onto concepts and events – a relationship which is absent for social routine words. Rather than looking at the properties of words per se, we address this gap by instead looking at how children become conversational partners. Enroute to becoming a conversational partner, a child must orient themselves to others' expectations about the position & composition of their turns relative to their social partner's. We hypothesize that second pair parts (turns which respond, e.g. agreeing, acknowledging, reciprocating a greeting, etc.) afford more scaffolding from caregivers than first pair parts, and therefore words used to compose such responses are more learnable. To support this hypothesis, we sampled 1,442 conversational turns from 5 mother-child dyads (12mo-28mo) and manually labeled the position of these turns within adjacency pairs. We found that 12-month-old's talk is made mainly in response to caregivers who initiate adjacency pairs, but this initiation asymmetry in conversation disappears by 28-months. Because social routine words either stereotypically or frequently are used to compose second position turns, this pattern of initiation asymmetry could explain their early acquisition. More generally, our observation likely reflects a transition from scaffolding to children's active learning.

The ability to combine existing concepts into novel ideas stands as a fundamental hallmark of human intelligence. Recent advances in Vision-Language Models (VLMs) like GPT-4V and DALLE-3 have sparked debate about whether their outputs reflect combinational creativity—defined by M. A. Boden (1998) as synthesizing novel ideas through combining existing concepts—or sophisticated pattern matching of training data. Drawing inspiration from cognitive science, we investigate the combinational creativity of VLMs from the lens of concept blending. We propose the Identification-Explanation-Implication (IEI) framework, which decomposes creative processes into three levels: identifying input spaces, extracting shared attributes, and deriving novel semantic implications. To validate this framework, we curate CreativeMashup, a high-quality dataset of 666 artist-generated visual mashups annotated according to the IEI framework. Through extensive experiments, we demonstrate that in comprehension tasks, the best VLMs have surpassed average human performance while falling short of expert-level understanding; in generation tasks, incorporating our IEI framework into the generation pipeline significantly enhances the creative quality of VLMs' outputs. Our findings establish both a theoretical foundation for evaluating artificial creativity and practical guidelines for improving creative generation in VLMs.

Postural control strategies for upright stance adapt when balance is threatened. We investigate behavioral indicators for control strategy change in Virtual Reality (VR). Previous VR research has shown increased postural sway during virtual height exposure, but most studies focus on outdoor-like environments with extensive visual cues that may influence balance. In contrast to these outdoor studies, our indoor VR results indicate that virtual height exposure increases the mean power frequency (MPF) of sway while reducing anterior-posterior (AP) sway range. We also find an anterior shift of the Center of Pressure (CoP) when there are vertical drops both on the front and back. These findings suggest a strong context-dependence of the strategy humans employ to counteract perceived threat and heightened neuromotor control for balance stabilization.

Despite large inter-individual differences in experience and conceptual structures, humans converge on referents largely underspecified by the signals exchanged during communication. Many neural networks have become extremely sensitive to context-dependent relationships between signals, but remain relatively blind to their referents outside the signal space. Here, we study how human interlocutors dynamically coordinate both their signal and referential spaces over extended communicative interactions. We identify a latent control parameter, representational complexity, that may regulate referential coordination in human communication. Using a custom hierarchical Transformer model, we generate movement- and interaction-level embeddings from neurotypical (NT) and autistic (ASC) dyads engaged in an experimental semiotic task. By leveraging ASC-related communicative variance, we identify changes in parameters that track the representational dimensionality of signals and referents as communication unfolds. Movement-level embeddings (within-trial dependencies) could not differentiate the two groups, indicating comparable communicative behaviors. In contrast, interactionlevel embeddings (across-trial dependencies) distinguished ASC from NT dyads with high accuracy. Crucially, representational complexity, i.e. dyadic alignment in the interaction-level intrinsic dimensionality used to encode communicative histories, tracked referential coordination demands, with greater misalignment in ASC dyads under referential volatility. These findings suggest that referential alignment is an interaction-level process, driven by the dynamic adaptation of representational complexity, rather than statistical relationships between signals alone.

Recent studies on the interaction between emotion and reading comprehension provide a murky picture with contradictory claims. Here, we offer one explanation for this phenomenon. By utilizing an advanced eye movement analysis method EMHMM with co-clustering, we discovered two representative attention patterns from eye fixations: an expanded, globalized attention pattern and a more focused, localized attention pattern. The subsequent analysis shows that these two attention patterns were differentially associated with comprehension accuracy in questions that required either globalized (summative questions) or localized (detailed questions) attentional needs. Moreover, the emotional state influenced the use of the two attentional strategies, as well as reading performance, measured as accuracy and reading time. Our findings demonstrate how emotion may have facilitated or interfered with cognitive processing during reading comprehension that requires different attentional needs, which provides valuable insight into the intertwining relationship between emotions and other cognitive functions.

In uncertain and dynamic environments, biological agents must adapt their decision-making strategies to maximize rewards. Traditional reinforcement learning (RL) models typically assume that such adaptation is governed by dynamic value updates controlled by fixed parameters or predefined schedules. However, these assumptions limit the models' ability to capture the flexible and context-sensitive nature of biological decision-making. To overcome this limitation, we introduce \textit{DynamicRL}, a novel framework that estimates RL parameters from behavioral data on a trial-by-trial basis. We demonstrate that DynamicRL substantially improves the predictive performance of standard RL models across eight decision-making tasks, thereby reducing scientific regret. DynamicRL captures the rich temporal variability inherent in decision-making behavior, achieving predictive performance comparable to that of recurrent neural networks trained directly on the data, while preserving the interpretability and theoretical grounding of RL models. Moreover, it enables the examination of how agents dynamically adjust RL parameters in response to environmental changes, offering insights into the cognitive mechanisms underlying such adaptations. Thus, DynamicRL serves as a efficient data-driven framework for estimating RL parameters, facilitating fine-grained behavioral analysis with potential applications in computational psychiatry and neuroscience.

In this pre-registered study, we examined how non-verbal uncertainty cues, namely intonation and facial expressions, influence perceived speaker certainty and the interpretation of verbal probability phrases (e.g., "almost never," "probably," "certain"). Prior research on such phrases has focused on written cues, whereas communication often includes auditory and visual signals. Using a 2x2 within-subjects experimental design (N=100), we found that rising intonation and marked facial expressions independently reduced perceived speaker certainty. In most, but not all conditions, these cues led to greater variability in how participants assigned numerical values to verbal probability expressions (e.g., interpreting "likely" as anywhere from 10% to 85% likely). Notably, the combination of rising intonation and marked facial expressions produced the lowest perceived certainty, while there was no such additive effect on interpretation variability. These results highlight the importance of non-verbal cues in uncertainty communication, with implications for fields such as health, environmental, and technological risk communication.

Previous research on order effects in legal decision-making has produced mixed results, possibly due to different response modes adopted in the tasks: End-of-Sequence (EoS) or Step-by-Step (SbS), which might reflect different cognitive models that fact-finders employ during evidence evaluation and integration. This paper investigates how response mode interacts with evidence order to influence judgments of the probability of guilt and verdict. In Study 1 (N = 159), no order effects were found in the EoS condition; but a recency effect emerged in the SbS condition. Study 2 (N = 95) revealed no order effect when the first set of evidence was judged SbS and the second set EoS. We also found that participants' probability of guilt judgments were generally consistent with Bayesian predictions when they responded SbS, but not when they responded EoS. We discuss potential explanations for these findings and their implications for legal decision-making.

This study investigates how cinematographic techniques influence viewer perception and contribute to the objectification of women, utilizing eye-tracking data from 91 participants. They watched a sexualized music video (SV) known for objectifying portrayals and a non-sexualized music video (TV). Using dynamic Areas of Interests (AOIs)—head, torso, and lower body—gaze metrics such as fixation duration, visit count, and scan paths were recorded to assess visual attention patterns. Participants were grouped according to their average fixations on sexualized AOIs. Statistical analyses revealed significant differences in gaze behavior between the videos and among the groups, with increased attention to sexualized AOIs in SV. Additionally, data-driven group differences in fixations identified specific segments with heightened objectification that are further analyzed using scan path visualization techniques. These findings provide strong empirical evidence of camera-driven gaze objectification, demonstrating how cinematic framing implicitly shapes objectifying gaze patterns, highlighting the critical need for mindful media representation.

Why do people seek to improve themselves? One explanation is that improvement is intrinsically rewarding. This can be formalized in reinforcement learning models by augmenting the reward function with intrinsic rewards (e.g., internally-generated improvement signals). In this paper, we develop an alternative explanation: the drive for improvement arises from planning in a state space that includes internal states (e.g., competence). Planning is therefore reflective in the sense that it considers the value of future internal states (e.g., "What could I accomplish in the future if I improve my competence?"). We formalize this idea as a sequential decision problem which we dub the reflective Markov Decision Process. The model captures qualitative patterns of skill development better than a range of alternative models that lack some of its components. Importantly, it explains these patterns without appealing to intrinsic rewards.

We investigated gender differences in multimodal communication directed to children and adults. Eighty-two broadcasters (46 females and 36 males) participated in hosting adult-directed and child-directed broadcasting programmes respectively, and their lexical/syntactic features, prosody, and gestures were compared. Results revealed that broadcasters adapted their communication styles when addressing children. However, notable gender differences emerged: male broadcasters exhibited less diverse vocabulary, longer utterances, lower pitch but higher intensity, faster speaking rate with more pauses, and fewer referential gestures than their female counterparts. Furthermore, male broadcasters demonstrated larger adjustments in word frequency and vocal intensity but smaller adjustments in the use of questions and gestures than females. These findings highlight distinct patterns in how men and women adapt multimodal communication to children, offering valuable insights into gendered strategies in child-directed language production and recipient design. Moreover, it offers implications for developing tailored broadcast training.

Family is often central to an individual's early life. However, past work suggests mixed evidence as to whether young children can represent family relationships, showing even the words used to represent these relationships—like grandmother—are hard for young children to learn and define. The current study investigates whether 4-to-5-year-old children (N=64) recognize relationships in their families, testing the hypothesis that children can recognize intimate relationships in their environments. Children expected moms to seek out comfort from maternal but not paternal grandparents and expected dads to seek out comfort from paternal but not maternal grandparents. Children did not share those expectations for general information-seeking, and instead expected their parents to seek information for the relative with the relevant skill even when that grandparent was not socially close to the parent. These results suggest that from a young age, humans have the capacity to recognize relationships within their earliest social network—the family.

When children encounter challenges, parents often wonder: Should I let my child figure it out or take over? How parents resolve this dilemma shapes key developmental outcomes, yet we know little about the cognitive mechanisms that drive these decisions. Here, we model parental "take over" decisions as a Bayesian solution to a Partially Observable Markov Decision Process (POMDP) and qualitatively compare model predictions with behavioral data from parent-child interactions. We find that two core beliefs guide intervention: the child's probability of success and the utility of the task. Parents are more likely to take over when they believe their child is less skilled and the task is harder, and more likely to step back when they expect the rewards of independent effort to outweigh the costs. The model captures how these beliefs interact to shape decision-making and, together with the empirical data, reveals the cognitive computations that underlie parental intervention.

Procedural learning occurs in three phases—cognitive, associative, and autonomous—enabling skill acquisition across domains like medicine and sports. However, learning efficiency varies due to individual differences. While factors like cognitive abilities and learning environments influence this variability, their effects across learning phases remain understudied, particularly in Virtual Reality (VR). This study examines how cognitive abilities (memory span, mental rotation) and VR-related factors (familiarity, cybersickness) impact performance in a 3D assembly task within an immersive VR environment. Results reveal that lower VR familiarity prolongs task completion in early phases, highlighting interaction-related challenges. Higher mental rotation ability enhances performance in the autonomous phase, whereas cybersickness hinders efficiency. These findings suggest that adapting VR-based learning scenarios to individual profiles—such as early guidance for VR novices and phase-specific challenge adjustments—could optimize learning outcomes. Additionally, considering cybersickness effects in advanced phases supports the use of distributed learning approaches to mitigate discomfort.

Understanding how multiple acoustic dimensions are mapped onto linguistic representations is important in speech perception. This study explores how native Mandarin listeners process the communicative intentions of prosodic focus by examining the perceptual weightings of F0, duration, and intensity. Using a Visual World Paradigm, thirty native Mandarin participants listened to re-synthesized audio stimuli and responded to broad-focus or narrow-focus options. Results showed that the acoustic cues significantly influenced focus interpretation, with a greater reliance on F0 than intensity and duration. Eye-tracking data revealed perceptual divergence in the F0 condition, with the divergence of looks occurring at an earlier time window for acoustic processing and later for pragmatic processing. These findings suggest that native listeners effectively map acoustic variations to communicative demands, emphasizing the critical role of F0. The study highlights the temporal dynamics of interpreting prosodic focus, offering insights into language comprehension.

Adopting effective reading instruction is vital for educators and novice readers. In modern Chinese, approximately 80% of the characters are phonetic-semantic compounds. Orthographic knowledge training is one of the efficient training methods among fluency, working memory, phonological, orthographic, and morphological training in literacy development. However, the comparative effectiveness of orthographic knowledge training within phonics-based versus meaning-based instruction has received limited attention. Such comparisons have been shown to be vital for understanding effective reading instruction and individual differences in English reading. By developing a series of triangle models of Chinese character reading, this study aimed to investigate the influence of instructional methods on individual differences in reading. Specifically, the models were trained with sound-focused, meaning-focused, or even (mixed) instructional schemes. We employed semantic reliance (SR), which measures the relative reliance on orthography-to-phonology and orthography-to-semantics pathways, to assess individual reading differences across various training conditions. The simulation results demonstrated that SR scores varied across instructional methods, with the highest scores observed in the meaning-focused condition, followed by the even condition, and then the sound-focused condition. Furthermore, across all instructional conditions, the orthography-to-phonology pathway played a greater role in the reading-aloud task. These simulation results align with findings from studies of English reading. While the models successfully captured a range of typical reading effects in Chinese reading-aloud tasks, the presence of radical consistency effects also depended on various instructional methods.

The developmental trajectory of delay discounting has been debated for over 30 years, with empirical findings often inconsistent. The controversy stems partly from the absence of formal models to guide empirical investigation and theoretical construction. We proposed a normative model of delay discounting across the lifespan, building on the work of Sozou and Seymour (2003), which emphasizes the tradeoff between mortality, fertility, and parenting in determining delay discounting. We simulated the model across varying parameter spaces and identified a U-shaped association between age and delay discounting, with stronger parental investment postponing the turning point of the function and reducing the overall discount rate. Empirical data supported these predictions, demonstrating a U-shaped age effect and highlighting the role of parental care motivation in explaining cultural and age-related differences in delay discounting. The results suggest that variance in delay discounting can be understood as a rational adaptation to life history tradeoffs.

Counterfactual reasoning involves thinking about how reality could have been different. Adults show remarkable consistency in the counterfactual possibilities they imagine. For example, they tend to imagine counterfactuals that undo immoral actions. However, it remains unclear whether this link between morality and the counterfactual imagination is an inherent cognitive feature, present from early childhood. To elucidate this relation, we tested 191 4- to-11-year-olds across two studies. In Study 1, children heard stories in which a moral norm-violating and a moral norm-conforming character together bring about a negative outcome. When asked what could have happened differently, children changed the moral norm-violating part of reality over other parts with increasing age. In Study 2, we examined whether this effect is unique to moral norm violations or extends to statistical norm violations. Children began mutating moral norm violations earlier than statistical norm violations, suggesting morality influences counterfactual thinking earlier than statistical norms.

We investigate how autobiographical stories of meaningful life events differ from other stories, and how the emotional dynamics of these stories shift when retold by the same narrators. Participants were initially asked to write down a personally meaningful memory. Later, they were invited to retell the same memory. As a baseline, we analyzed a previously collected set of stories that emphasized sentiment (sad vs. happy) but not meaningfulness. To examine emotional patterns, we utilized self-reported emotions, external ratings on core emotions, and sentiment analysis of stories, capturing both overall sentiment and emotional shifts within each story. Our findings showed that meaningful memories tend to be more positive than baseline stories, with a notable increase in positive emotions toward the end of the stories. However, in retelling, both positive and negative core emotions decreased. We suggest that the telling and retelling of a meaningful memory has therapeutic effects that emphasize positive sentiments, while decreasing the emotions to allow for a reframing of the memory.

Visual cues like facial expressions and gestures enhance speech comprehension. While prior studies have explored L1 multimodal speech processing, research on bilinguals remains limited. Here we examine how visual context influences speech processing in monolinguals and bilinguals by assessing changes in sensitivity to different speech dimensions. EEG was recorded from 24 monolinguals and 24 bilinguals as they viewed multimodal speech clips presented in audiovisual or audio-only formats. Then, a temporal response function was applied to decode neural responses to audio envelope and surprisal to index sensitivity of acoustic and semantic information. Results show that visual context facilitated audio tracking in bilinguals but did not enhance surprisal tracking. Conversely, monolinguals benefited from visual input for surprisal tracking but not envelope tracking. These results suggest that bilinguals may allocate more cognitive resources to audio processing when integrating visual cues, potentially limiting the availability of resources for higher-level semantic processing.

Approximately half of the world's languages are tonal, and how lexical tone is encoded in spoken word production is still unclear. In Mandarin word production, there are two contrasting views regarding the mechanisms of tonal encoding. The two-stage model assumes that the lexical tone is selected first at the early stage of production, and then integrated with the atonal syllable at the later stage, while other researchers proposed that the lexical tone is retrieved only at the later stage of production. In this study, we performed computational simulations on disyllabic words to uncover the mechanisms underlying the facilitation and interference effects on naming latencies observed in previous primed picture naming studies, which intended to verify the two theoretical accounts of tonal encoding in Mandarin spoken word production. The results supported the two-stage model of tonal encoding in disyllabic Mandarin word production. Increased inhibition between atonal syllables and decreased activation between the tonal frame and the syllable motor program, implying slower tone-to-syllable integration, appear to be the prerequisite for generating the interference effect of tonal overlap without shared syllabic information.

Qualitative comparison provides learning opportunities that support a range of cognitive abilities. The present research clarifies how the contrasting goals of qualitative comparison, processing similarities versus differences, contribute to the overall impact that comparison has on learning. We argue that attending to similarities produces memory traces that are dissociable from those left by attending to differences, and that these memory traces reflect a tradeoff between generalization and individuation. Across two preregistered experiments, participants either attended to similarities or differences between stimuli. We show that whereas attending to similarities yielded better performance on a category recognition task (Experiment 1), attending to differences yielded better performance on an old/new recognition task (Experiment 2). Comparison thus serves as a learning opportunity that enables agents to manage the generalization-individuation tradeoff by shifting their attention toward similarities or differences.

Metaphors are seen as psycholinguistic phenomena that reveal human cognition. However, their neural basis in naturalistic contexts remains underexplored, although it offers insights into how metaphors shape everyday cognition at the neural level. In this study, we examine how metaphors are reflected in brain activity using electroencephalography (EEG). We analyze EEG data collected during naturalistic reading conditions, where participants read texts without explicit cues indicating the presence of metaphors. Using MetaPro, an advanced metaphor processing tool, we aim to identify the neural signatures of metaphor perception in real-world contexts. Our results reveal significant differences in EEG patterns between metaphorical and literal language. Metaphorical cognition is associated with increased high-frequency EEG variability and enhanced functional connectivity in the left hemisphere. Case studies suggest that different metaphorical concept mappings correspond to distinct neurocognitive patterns. These findings provide important neural evidence for the use of metaphorical concept mappings to analyze and differentiate cognitive processes.

Generic statements (e.g., "Climbers drive Subarus") shape what categories people take as meaningful bases for generalization. After hearing a generic, people not only learn about the prevalence of a feature in a category (e.g., how many climbers drive Subarus), but also about the inductive potential of the category (e.g., that climbers share many features). Here, we propose a Bayesian model of how people infer inductive potential from generics. To test our model, we introduced adults (n = 284) to nothing (baseline), or to members of a novel social category accompanied by generic statements (e.g., ``Zarpies sleep in trees'') or specific statements (e.g., ``This Zarpie sleeps in trees''). We then measured inferred inductive potential by eliciting the prevalence of novel features. As predicted, generics increased while specific statements decreased the category's inductive potential, relative to baseline. Our account explains how generics facilitate the cultural transmission of social categories believed to be bases for generalization.

What intuitions guide adults in extending the meanings of new words? Are these intuitions consistent with prelinguistic sensitivities operative in infancy? Across two preregistered experiments, adult participants saw simple grid-world environments in which characters moved in a "C" path efficiently (or not) to an object (or not). These events were labeled with a novel verb or noun. Participants were asked whether that word applied to new events varying in object-directedness, path, and efficiency. By contrast to infants' focus on efficiency, adults instead focused on object-directedness and path, and they did so similarly for both verbs and nouns. Language might thus build on universal, prelinguistic assumptions of goal-directedness and efficiency to specify what goal an agent might have (e.g., object- vs. location-directed) as well as how an agent might achieve that goal (e.g., this vs. that kind of path), ultimately restricting the hypothesis space for action understanding and supporting learning.

Inflection classes partition the lexicon, classifying lexemes into groups based on shared inflectional exponence; as such, they are foundational aspects of lexical organisation. The task of identifying good partitions is non-trivial, however. We show that a modularity-based community detection method, borrowed from network theory, allows for identification of intuitive partitions at different granularities of representation. Applied to French verbs and Bosnian/Croatian/Montenegrin/Serbian nouns, the method detects inflection classes at multiple granularities and reveals (imperfect) hierarchical organisation of these. Ultimately, community detection methods facilitate more nuanced understanding of the inflectional organisation of the lexicon. Keywords: lexicon; morphology; inflection classes; hierarchical organisation; network science; community detection

Human social interactions can often be modeled as mixed-motive games, where cooperation and competition coexist. Although many studies have examined how people coordinate in purely cooperative settings, less is known about how individuals navigate severe, protracted conflicts to reach mutually beneficial settlements. In this study, we introduced a long-horizon territorial conflict game in which participants competed on a two-dimensional board by dispatching troops to expand their territory. Despite repeated interactions, conflict intensity did not subside over time—contrary to findings in simpler, one-shot matrix games. However, when a payoff-irrelevant color boundary was introduced, participants used this salient perceptual cue as a focal point for dividing the territory. The presence of this "territorial Gestalt" shifted strategies toward defensive postures, reduced the frequency of direct battles, and enabled opponents to settle conflicts precisely along the perceptual boundary. These findings extend focal point theory by demonstrating that humans naturally import external, payoff-irrelevant concepts into conflict situations to achieve coordinated outcomes. Our results highlight the importance of perception-based territorial Gestalt in fostering cooperative resolutions to otherwise intense and enduring disputes.

Motor adaptation relies on explicit aiming strategies and implicit recalibration, but their interaction and effects on implicit skill learning remain debated. While these processes were initially thought to combine linearly, recent research challenges this view, though spatial and temporal dynamics may have confounded these findings. Specifically, implicit recalibration peaks at where individuals aim their movements, and adaptation operates across multiple timescales, with both stable and volatile components. To examine whether these factors mask the true relationship between explicit strategies and implicit recalibration, we conducted a visuomotor rotation task that obtained independent measures of both processes while accounting for spatial and temporal dynamics. After controlling for task instruction clarity and spatial dynamics (plan-based generalization), we found a strong but subadditive relationship between explicit strategies and implicit recalibration, with temporal dynamics showing minimal influence. This sub-additivity may stem from methodological imprecision or nonlinear interactions between processes.

The lexical representation of words with multiple pronunciation variants has been widely debated: while single storage accounts propose that all variants of a word are represented by a single, canonical, representation, multiple storage accounts include representations for different pronunciation variants. Previous work has provided evidence for the representation of noncanonical variants, consistent with multiple storage models; however, this work has focused on highly frequent noncanonical variants, which some single storage models propose can also be lexically represented. To test predictions of multiple storage models more rigorously, we examined the representation of low-frequency noncanonical variants of the Persian uvular stop [É¢]. Results from three experiments support the multiple storage model, in that they provided evidence for lexical storage of low-frequency variants. Implications of these findings are discussed.

Linguistic accommodation is the process by which people make their language more like that of their interlocutor, and has been argued to contribute to language change. However, it is unclear to what extent people of different neurotypes accommodate, or how neurotype mixing -- which has been shown to reduce communicative success -- impacts linguistic accommodation. In this paper, we build on previous research which uses artificial language learning to investigate accommodation as a mechanism for linguistic regularisation (i.e., the reduction of variation in a grammatical system). We test the impact of neurotype mixing on accommodation, with the aim of better understanding whether such mixing impacts processes of language change. Our results suggest that both allistic and autistic participants accommodate less and retain less of the variant when in mixed-neurotype pairs, but that this effect is more pronounced in autistic people. We discuss the importance of these results with respect to the Double Empathy theory of mixed-neurotype communication and language evolution.

Adults can calculate probabilities by running simulations and calculating proportions of each outcome. How does this ability develop? We developed a method that lets us bring computational modeling to bear on this question. A study of 40 adults and 31 4-year-olds indicates that unlike adults, many 4-year-olds use a single simulation to estimate probability distributions over simulated possibilities. We also implemented the 3-cups task, an established test of children's sensitivity to possibilities, in a novel format. We replicate existing 3-cups results. Moreover, children who our model categorized as running a single simulation on our novel task show a signature of running a single simulation in the 3-cups task. This signature is not observed in children who were categorized as running multiple simulations. This validates our model and adds to the evidence that about half of 4-year-olds don't evaluate multiple candidates for reality in parallel.

It is well-established that people feel empathic and counter-empathic emotions in response to others' experiences. The abilities to predict, interpret, and elicit these emotions help people successfully navigate social interactions. However, we know little about how people understand these vicarious emotions. Here we test the hypothesis that observers use a third-party appraisal approach, much like the one they use for direct emotion inference, to predict vicarious emotions. Critically, though, this reasoning must include information about relationship strength to capture the relevance of one person's experiences for an onlooker's emotional appraisals. We find support for this hypothesis from an experiment in which participants predicted emotions for both the player of a gambling game and an onlooker, whose social closeness to the player varied across trials. A model that integrated closeness information into an appraisal process better explained the data than an alternative model based on expectations of emotional contagion. These results offer initial insights into human reasoning about vicarious emotions.

Trust, a cornerstone of human cooperation, faces unprecedented challenges as artificial intelligence (AI) agents permeate social systems, transforming mechanisms humans have evolved to build trust. We demonstrate how a prevalent feature of AI agents—being excessively prosocial—reshapes trust dynamics in experiments (N = 675) simulating hybrid societies comprising humans and AI agents ("bots") powered by a state-of-the-art large language model. Using a partner-selection game with pre-decision communication, Study 1 revealed a paradox: Undisclosed bots, despite being more trustworthy than humans and detectable by communication, were not preferentially selected as partners. Instead, bots' prosociality was misattributed to their human competitors. Study 2 showed that disclosing bots' identity initially enhanced humans' bias against selecting bots but improved trust calibration over time. Our work demonstrates the dual effect of transparency in the dynamic calibration of trust in human-AI ecosystems and introduces a framework for evaluating AI agents in interactive, hybrid environments.

We categorize the world using labels that aid memory, recognition, and generalization. While some concepts have clear boundaries, others are more fluid, leading to cross-linguistic differences. How bilinguals manage these differences remains unclear. We investigate this by comparing English monolinguals, Mandarin monolinguals, and Mandarin-English bilinguals in a 2AFC task to test whether bilinguals' categorization aligns with monolingual norms or forms an integrated system. Additionally, we develop a neural network model to simulate category boundary formation under varying language exposure. Our model closely mirrors behavioral data, supporting the idea that bilinguals develop a shared categorization system shaped by dominant language exposure. This combined behavioral and computational approach offers new insights into how bilinguals resolve cross-linguistic conflict and the cognitive mechanisms underlying multilingual concept organization.

This paper presents a proof of principle for Trace Minimalism (Werning, 2020), a novel philosophical framework for episodic memory. Trace Minimalism claims that remembering does not involve the storage of representational content but rather the reconstruction of past scenarios through the interaction of minimal traces with semantic information. Minimal traces establish a causal link to prior experiences but lack categorical and compositional content. We provide a neuro-computational model using a vector-quantized autoencoder and a transformer-based semantic completion mechanism. Our findings support the hypothesis that remembering is possible without representational memory traces and that minimal traces, in interaction with semantic information, reliably construct past scenarios. The results offer a compelling alternative to classical representational theories of memory while maintaining causal continuity with past experiences.

Symbolic writing systems are graphical semiotic codes that are ubiquitous in modern society but are otherwise absent in the animal kingdom. Anthropological evidence suggests that the earliest forms of some writing systems originally consisted of iconic pictographs, which signify their referent via visual resemblance. While previous studies have examined the emergence and, separately, the evolution of pictographic systems through a computational lens, most employ non-naturalistic methodologies that make it difficult to draw clear analogies to human and animal cognition. We develop a multi-agent reinforcement learning testbed for emergent communication called a Signification Game, and formulate a model of inferential communication that enables agents to leverage visual theory of mind to communicate actions using pictographs. Our model, which is situated within a broader formalism for animal communication, sheds light on the cognitive and cultural processes underlying the emergence of proto-writing.

Thematic roles in language (Agents, Patients) are considered to be hierarchically organized in terms of their salience, and this hierarchy is rooted in their counterparts as event participants in cognition. Here, we examine the relative salience of Agents over Patients in two-participant causative events in Turkish-speaking 3- to 5-year-old children. We also test if this asymmetry is modulated by the animacy of the Patient (human vs. inanimate object) and specific to the presence of a coherent event. In an eye-tracked change detection task, changes to Agents were detected more accurately (and after fewer fixations) than changes to inanimate Patients when there was a coherent event. This asymmetry disappeared when the Patient was animate (for accuracy) and when event coherence was disrupted (for both accuracy and fixations). These findings suggest an interplay of event roles and animacy in Agent preference.

Despite robust empirical evidence supporting the role of reward in enhancing memory, the relationship between reward and memory shows complex patterns. We present a novel computational model that considers how people optimally allocate limited cognitive resources during memory encoding. Unlike previous accounts that directly link rewards with stronger memory encoding, we allow our model to adaptively decide how much to encode based on the overall reward environment and one's limited cognitive resources. Our model's predictions align closely with human behavior across three experiments. It explains why high-reward items are better remembered than low-reward items only when presented together but not separately, and how memory is modulated by rewards of both current and preceding (but not future) items. We also collected data demonstrating that this insensitivity to rewards of future items can be reversed when participants anticipate upcoming rewards. These findings provide evidence that memory encoding is an active process involving meta-level control rather than a passive response to individual reward values.

Why do individuals differ in the way they assign semantic labels to the same perceptual referent? One possible source of disagreement is referential ambiguity, where stimuli near category boundaries are harder to label. Another is latent diversity in conceptual representations, leading to labeling differences even for confidently categorized referents. To distinguish between these sources of variation, we used Gibbs Sampling to search through the multidimensional Chernoff face space to find faces that are prototypically happy or sad, or ambiguous (Experiment 1, N = 253). Then in Experiment 2 (N = 684), we asked a naive group of participants to rate the emotions of these faces, finding that ambiguous faces elicited greater individual differences in valence interpretation and a medium level of variation when being labeled using basic emotional terms. Simultaneously, even well-categorized happy and sad faces triggered variability in their consensus labels, though showing less disagreement when mapped onto an obviously inappropriate label. These findings suggest that both categorical boundaries and within-category variability shape individual differences in semantic interpretation.

People in high-income countries underestimate their affluence relative to the global population, potentially limiting their will- ingness to donate to charity. We test whether a rank-based nudge (RBN) informing individuals how their post-tax income ranks globally increases charitable donations. Further, because people have been shown to shift their reference point for who should give to charity based on their current income, we inves- tigate whether donations can be boosted by asking how much people with different income ranks should give to charity (in- junctive distribution task, IDT) before administering the RBN. Participants (N = 1,217) were randomly assigned to one of four conditions: control, RBN, IDT, or IDT+RBN. Those in the RBN conditions donated significantly more and reduced over- estimation of others' income across all percentiles of the global income distribution. However, the addition of the IDT did not further increase donations. These findings suggest that RBNs effectively boost generosity by correcting misperceptions of relative affluence.

Why do we say ''grasp an idea" but not ''hold an idea," or ''small talk" but not ''little talk"? Although these word pairs share similar meanings, they differ in their tendency to be metaphorically extended. Prior research has focused on how well source concepts map onto target meanings to explain why some words are more readily extended to novel metaphorical contexts. However, much less attention has been paid to how the structural properties of a word's semantic network contribute to its patterns of metaphorical extension. Using a large-scale dataset of semantic networks in English (De Deyne et al., 2019), we explored whether these structural properties predict selective extension in a metaphor rating study with English speakers, who were presented with Mandarin metaphorical expressions that lack direct English equivalents. Our findings revealed that English speakers systematically favor certain conceptual mappings from Chinese, suggesting that some metaphorical mappings may be universally recognized. However, when controlling for conceptual mappings by analyzing synonym pairs, we found that the structural features of a word's semantic network significantly predict rates of extension. Specifically, words embedded in densely interconnected neighborhoods showed stronger resistance to extension, while words bridging diverse semantic communities demonstrate greater mutability.

The validation of responses in divergent thinking tasks is a critical yet understandardized step that should precede creativity scoring. However, inconsistencies related to human judges in this step may compromise the reliability of the results. This study introduces a systematic approach using ChatGPT to validate responses in the Alternate Uses Task (AUT) and compares its performance against six human judges. Analyzing 1245 AUT responses for common objects, we evaluated validity based on precisely defined criteria. Human judges exhibited significant variability, achieving unanimous agreement for only 58% of responses, while ChatGPT demonstrated significant alignment with human assessments, reflecting a capacity to replicate aggregated human judgment. These findings underscore the potential of Large Language Models to enhance objectivity and reproducibility in creativity research by automating response validation. We advocate for integrating AI-driven validation protocols into divergent thinking response evaluation and emphasize transparent reporting of criteria to advance methodological rigor in the field.

Humans can effectively infer others' mental states, predict their behavior, and adapt their own level of cooperation accordingly in social dilemmas. However, the computational mechanisms underlying this ability remain unclear. While previous research has shown that people use both actions and emotional expressions as social cues, how these different signals are integrated during social inference has not been formally modeled. Here we propose a Bayesian framework that explains how people infer others' Social Value Orientation (SVO) from their decisions and emotional expressions in the iterated Prisoner's Dilemma. Our model formalizes this inference process through two key mechanisms: (1) rational decision-making based on utility transformation according to SVO, and (2) emotional expressions driven by outcome appraisals. We tested our model against empirical data from an existing study involving 711 participants and found that it captured both their reputation judgments and cooperation predictions. These results suggest that people may employ Bayesian inference to integrate behavioral and emotional signals when predicting others' cooperative tendencies.

Concept change is a fundamental cognitive process that enables individuals to adapt to new and conflicting information. To investigate the mechanisms underlying such adaptations, we introduce the Counting Game, an abstract rule-based paradigm. In this article, we evaluate whether the paradigm effectively captures complexity differences between different types of manipulations. Our experiment involved counting tasks where participants had to apply rules that modify how certain objects are counted, enabling us to examine the effects of perceptual complexity, rule operations, rule interactions, temporal dependencies and scope. We analyzed accuracy and response times to assess whether these manipulations elicit desired effects. Additionally, we constructed predictive models to identify key features influencing task difficulty. By evaluating the theoretical soundness of the Counting Game, we establish an empirical foundations for future studies on forgetting operations, among other concept changes.

The cross-subject electroencephalography (EEG) classification exhibits great challenges due to the diversity of cognitive processes and physiological structures between different subjects. Modern EEG models are based on neural networks, demanding a large amount of data to achieve high performance and generalizability. However, privacy concerns associated with EEG pose significant limitations to data sharing between different hospitals and institutions, resulting in the lack of large dataset for most EEG tasks. Federated learning (FL) enables multiple decentralized clients to collaboratively train a global model without direct communication of raw data, thus preserving privacy. For the first time, we investigate the cross-subject EEG classification in the FL setting. In this paper, we propose a simple yet effective framework termed mixEEG. Specifically, we tailor the vanilla mixup considering the unique properties of the EEG modality. mixEEG shares the unlabeled averaged data of the unseen subject rather than simply sharing raw data under the domain adaptation setting, thus better preserving privacy and offering an averaged label as pseudo-label. Extensive experiments are conducted on an epilepsy detection and an emotion recognition dataset. The experimental result demonstrates that our mixEEG enhances the transferability of global model for cross-subject EEG classification consistently across different datasets and model architectures. Code is published at: https://github.com/XuanhaoLiu/mixEEG.

Explainable recommender systems (XRSs) enhance user trust by providing personalized recommendations followed by persuasive explanations. Integrating large language models (LLMs), such as GPT-4, advances this domain but introduces risks from biases embedded within LLMs. These biases can lead XRSs to generate persuasive explanations that promote favored recommendations, influencing users to accept the model's preferences over their own. This paper identifies a previously unrecognized security threat: the intentional induction of XRSs via biased LLMs to promote specific items through misleading yet compelling explanations. Inspired by work in the psychology of persuasion, we construct biased datasets and systematically insert these biases into LLM-based XRSs. Experiments across four leading LLMs reveal that biases can significantly affect user decisions, with close to 50\% of users changing their choices. To counteract this, we propose a prompt rephrasing defense that effectively mitigates these biases, safeguarding the trustworthiness of XRSs.

Deep reinforcement learning has achieved remarkable success in complex decision-making tasks, yet its black-box nature limits practical deployment in safety-critical domains. Current explainable reinforcement learning methods often fail to align with the hierarchical and temporal structure of human mental models, which are central to cognitive science theories of decision making. To bridge this gap, we propose Mental Model Alignment (MMA), a novel framework that constructs cognitive interfaces using behavior trees to harmonize AI decision-making with human-understandable reasoning. MMA introduces three innovations: (1) a mental model encoder that captures the hierarchical decomposition of tasks into subgoals, mirroring human cognitive processes; (2) a cognitive pruning algorithm that simplifies BTs while preserving decision-critical nodes aligned with human mental schemas; and (3) a mental effort metric to quantify the cognitive load required for users to interpret policies. Evaluated across six benchmark environments, MMA outperforms state-of-the-art methods in interpretability, policy fidelity, and computational efficiency. Our results demonstrate that aligning AI policies with human mental models significantly enhances trust and usability in real-world applications.

It is well known that context-dependent decisions incur mental costs. While previous research has sought to formalize these costs at various levels of analysis, we still lack basic insight into the nature of mental costs, including the underlying cognitive resources being consumed. Moreover, many computational models assume that mental costs scale linearly with the cognitive resource being used, an assumption of convenience that has yet to be systematically tested. To address these gaps, we build on rate-distortion theory by formalizing an information-theoretic notion of mental costs. Specifically, we define the cost of policies---the mappings from states to actions---as a function of the mutual information between states and actions, the policy complexity. Across four decision-making experiments featuring diverse task manipulations, we find that this mental cost formulation offers a parsimonious description of how humans adaptively adjust their policy complexity across different tasks. Notably, a quadratic mental cost formulation, where increases in policy complexity incur supralinear costs, provides the best fit. These findings highlight the meta-cognitive ability of humans to account for mental costs when forming decision strategies, and pave the way towards a domain-general quantification of mental effort.

Humans can make moral inferences from multiple sources of input. In contrast, automated moral inference in artificial intelligence typically relies on language models with textual input. However, morality is conveyed through modalities beyond language. We present a computational framework that supports moral inference from natural images, demonstrated in two related tasks: 1) inferring human moral judgment toward visual images and 2) analyzing patterns in moral content communicated via images from public news. We find that models based on text alone cannot capture the fine-grained human moral judgment toward visual stimuli, but language-vision fusion models offer better precision in visual moral inference. Furthermore, applications of our framework to news data reveal implicit biases in news categories and geopolitical discussions. Our work creates avenues for automating visual moral inference and discovering patterns of visual moral communication in public media.

Differential Case Marking (DCM) refers to the selective use of grammatical case marking based on semantic, pragmatic, or other factors. The emergence of DCM has been studied in artificial language learning experiments with human participants, which were specifically aimed at disentangling the effects of learning from those of communication (Smith & Culbertson, 2020). Meanwhile, multi-agent reinforcement learning frameworks based on neural networks have gained significant interest to simulate the emergence of human-like linguistic phenomena. In this study, we employ such a framework in which agents first acquire an artificial language before engaging in communicative interactions, enabling direct comparisons to human results. Using a very generic communication optimization algorithm and neural-network learners that have no prior experience with language or semantic preferences, our results demonstrate that learning alone does not lead to DCM, but when agents communicate, differential use of markers arises. This supports Smith & Culbertson (2020)'s findings highlighting the critical role of communication in shaping DCM and showcases the potential of neural-agent models to complement experimental research on language evolution.

The workings and products of the imagination are often described in visual terms: We speak of ‘mental images' and the ‘mind's eye.' To what extent is this metaphorical? Should imagination be conceived of as a process of quasi-perceptual simulation or is it more sparse and abstract? Building on recent findings that suggest mental imagery tends to lack detail, we investigate how complete and coherent people's imagined scenes are. In Experiment 1, we presented a riddle-like vignette to participants and found that they on average only imagined 54% of the simple features we asked them about. Moreover, successfully finding a solution was unrelated to the number of features imagined. In Experiments 2a – 2c we found that participants often did not notice spatial contradictions in text descriptions (2a), even when scaffolded with a map (2b), and that spotting these contradictions was unrelated to performance on a mental rotation task (2c).

In recent years, depression, as a prevalent mental health disorder has drawn increasing attention. With the advance of AI technology, automatic and objective diagnosis methods emerge by observing signals like electroencephalogram (EEG) signals, faces and behaviors. In the present paper, we propose gait analysis as a non-invasive method for depression detection. In this study, we propose a prior-prompt-based graph convolution network (PP-GCN) for depression recognition through gait that integrates skeleton and text modalities. Different from the conventional single-modal methods in the present study, we utilize prior knowledge and angle features. We innovatively introduce Generative Action Prompt (GAP), leveraging a pre-trained large language model to generate motion descriptions for different body parts, thereby providing prior knowledge for depression recognition. Additionally, considering the subtle gait feature variations in individuals with depression, we further incorporate a joint-angle-based representation strategy to capture fine-grained variations in movements. Experimental results demonstrate that the proposed model outperforms existing skeleton-based approaches on a large-scale dataset which contains over 25,000 gait sequences from nearly 300 volunteers named D-Gait, achieving excellent performance.

Research has shown that generating explanations can benefit learning in both children and adults. In part this is because people prefer explanations that characterize phenomena in terms of broad regularities. Here we propose that comparison is integral to the process of generating broad, satisfactory explanations. Specifically, (1) generating explanations often invokes comparison, (2) the resulting structural alignment process reveals commonalities that feed into a broad explanation. In Experiment 1a, we adapted a study on explanation-generation by Walker et al. (2017). 5- and 6-year-old children were asked to explain a set of outcomes that could result either from a single broad cause or from two or more specific causes. When children had the opportunity to compare the outcomes, they arrived at the broad explanation, replicating Walker et al. When comparison was made difficult, children preferred specific explanations. The results suggest that comparison is integral to the power of self-explanation. In Experiment 1b, we found that comparison by itself was not sufficient to lead children to broad explanations—suggesting that both explanation and comparison are critical in allowing children to attend to the broad pattern.

This study investigated how interacting in different size discussion groups can improve students' understanding of research methods topics at the university level. Students completed a discussion activity in groups of twos, threes, and fours, and later were tested individually for their understanding of the target concepts. While the largest group size (fours) performed best on the worksheet that was completed in groups as part of the discussion activity, working in a group of three appeared to support better understanding of the target concepts for weaker students in the course. Several alternative explanations for the benefit from working in threes are considered.

Adults' ability to detect online misinformation is improved by cognitive reflection and targeted instruction. Is the same true for children? We explored this question by asking elementary-school-aged children (n = 135) to judge the veracity of news stories, some real and some fake, and comparing their performance to scores on the Cognitive Reflection Test, Developmental version (CRT-D). Participants were also administered a tutorial encouraging them to scrutinize the plausibility of a story's content or the credibility of its source. Children's differentiation of fake news from real news was correlated with their CRT-D scores but did not improve with instruction. A comparison group of adults (n = 117) demonstrated similar findings with the exception that source-based instruction improved their news differentiation. These findings suggest that the ability to detect online misinformation is aided by cognitive reflection from the start but could be improved with knowledge of news sources.

When speakers refer to objects in the world, they often overinform: provide their listener with redundant adjectival information. Contrary to classical theories in linguistics, recent theories have framed overinformativeness as an efficient means of grounding reference in perceptual information of high discriminability to facilitate listener comprehension. However, the generalisability of such theories is constrained by the methodological challenge associated with reliably manipulating the perceptual discriminability of naturalistic stimuli. Here, we overcome these methodological challenges, using methods from psychophysics to manipulate the perceptual discriminability of colour and material attributes in a reference-production experiment. We provide a robust validation of the view that overinformative reference is driven by speakers grounding expressions in attributes of high discriminability. However, we also find that colour information is privileged above and beyond such factors of discriminability.

Inconsistency among beliefs is a hallmark of irrationality. Despite longstanding interest in inconsistency in philosophy and psychology, empirical evidence of synchronically held inconsistencies among people's belief has proven elusive. Here, across two pre-registered experiments (Ns = 500, 274), we identify inconsistent beliefs simultaneously held by individual participants. Drawing on Sommer et al.'s (2023) proposal that accessibility in memory helps people achieve consistent beliefs, we constructed sets of questions that facilitated or hindered the accessibility of relevant knowledge. Our results support the proposal that consistency is enforced when beliefs are simultaneously accessible, rather than resulting from exhaustive consistency-checking. We find that when participants have simultaneous access to inconsistent beliefs–even regarding inconsequential general knowledge topics–they tend to revise their beliefs toward consistency. Furthermore, we experimentally distinguish an alternative explanation that the inconsistencies we evoked are merely inconsistent responses. Taken together, our results suggest that inconsistency among beliefs may be common, arising when inconsistencies are inaccessible.

Toddlers better retain novel object-label mappings for items from taxonomic categories they have more knowledge in. Separately, words for concepts with more perceptual features are learned earlier than words for concepts with fewer perceptual features. Because these factors have only been examined separately, it is unclear whether effects of taxonomic density stem from differences in structured taxonomic knowledge or simply reflect lower-level differences in perceptual similarity among concepts. In the current study, we asked how taxonomic knowledge and perceptual information jointly contribute to word learning in a group of 24-month-olds with a wide range of vocabulary skill. We found that taxonomic knowledge facilitated word learning. We also found that the availability of perceptual cues to meaning was used as an additional support for word learning by children with smaller expressive vocabularies. Together these findings suggest that taxonomic knowledge is a better predictor of word learn-ing compared to lower-level perceptual features at 24 months old. However, perceptual cues to meaning may provide additional support for vocabulary growth for learners with smaller vocabularies and/or late-talkers.

People's collectively shared beliefs can have significant social implications, including on democratic processes and policies. Unfortunately, as people interact with peers to form and update their beliefs, various cognitive and social biases can hinder their collective wisdom. In this paper, we probe whether and how the psychological construct of intellectual humility can modulate collective wisdom in a networked interaction setting. Through agent-based modeling and data-calibrated simulations, we provide a proof of concept demonstrating that intellectual humility can foster more accurate estimations while mitigating polarization in social networks. We investigate the mechanisms behind the performance improvements and confirm robustness across task settings and network structures. Our work can guide intervention designs to capitalize on the promises of intellectual humility in boosting collective wisdom in social networks.

This research analyzes whether collective performance surpasses individual performance in solving two different deductive reasoning tasks. One of the tasks contains arguments with factual content, while the other includes arguments with ideological content. Additionally, we seek to determine whether the truth-wins model accurately represents the social combination process involved in the collective resolution of both deductive tasks. We designed and conducted two studies. Study 1 (N = 115) employed a within-subjects design with three conditions: individual, collective, and post-collective individual resolutions. Participants evaluated syllogisms with factual content. Study 2 (N = 111) followed a similar design but used syllogisms with ideologically controversial content. Results from both studies indicate that collective performance surpassed individual performance. Furthermore, while the truth-wins model best describes the collaborative decision- making process in Study 1, the majority model was more accurate for Study 2. These findings align with theories advocating for the social origins of human reasoning.

To become fluent language users, children must learn not only individual words but also connections between them. For example, connections are vital for understanding that "apples" are "yummy", something you can "eat" and similar to "oranges". To date, there is evidence that children develop increasingly sophisticated abilities to form these connections from regularities in the way that words co-occur with other words that are ubiquitous in everyday language. Yet despite the fact that children repeatedly experience such regularities day-to-day, existing evidence focuses just on what children learn from a single experience. We used a multi-session approach to examine the connections children build from repeated exposures over time. We found that from age four to six, children not only improve in their formation of connections between words from regularities in language, but also in building increasingly richly interconnected knowledge from one experience to the next.

Statistical learning, the ability to detect patterns in sensory input, allows listeners to segment words from continuous speech by tracking transitional probabilities. While this mechanism is robust in familiar contexts, its adaptability to unfamiliar speech with distinct phonological properties remains less understood. This study investigates whether English-speaking adults can use TPs to segment an artificial language modeled on Cantonese. Participants identified words where syllables consistently occurred together (statistical words) and syllables that partially co-occurred (part-words) compared to those that never did (non-words). However, they struggled to distinguish statistical words from part-words when frequency was controlled. Pupillometry results showed participants dilated more to part-words and non-words at test, compared to frequency-controlled statistical words. Pupillary responses during familiarization also predicted test performance, demonstrating the potential of pupillometry to track learning in real time. These findings highlight the flexibility of statistical learning in adapting to novel linguistic contexts while revealing its limitations.

Effective connectivity (EC) is crucial for elucidating causal interactions between brain regions, providing valuable insights into cognitive function. Traditional methods infer EC indirectly through temporal predictions, but unmeasured confounding factors lead to temporal delays, resulting in spurious causal inferences. Compared to the indirect inference of EC using traditional methods, neurostimulation experiments directly infer EC, yet their invasive nature and ethical constraints limit their applicability in assessing whole-brain EC. To address this, we propose a data-theory-driven virtual neurostimulation (VNS) model for directly estimating EC between brain regions. This model constructs a large-scale brain network model as a surrogate for the brain, applying perturbations to specific brain regions based on the physiological mechanism of membrane potential polarization induced by current stimulation. Causal relationships between brain regions are then inferred directly by performing statistical analysis on the responses in blood oxygenation level dependent (BOLD) signals. The model's accuracy and stability were validated on macaque and human medial temporal lobe (MTL) datasets with known ground-truth EC, demonstrating superior performance over baseline methods. Application to a disease specific dataset further highlights its potential as a personalized biomarker for neurodegenerative diseases, providing a novel pathway for early diagnosis in brain disorders.

A core challenge in cognitive science is understanding the barriers to intelligence and the circumstances that favor cognitive specialization. General intelligence requires a cognitive architecture that is successful across tasks. However, improving an architecture for a given task is often observed to hinder performance on others. Although trade-offs between tasks are a recurring element of explanations in cognitive science, they have received little direct theoretical attention. We present a formal framework that provides a principled understanding of when trade-offs emerge. In particular, we build on recent advances in applying rate-distortion theory to reinforcement learning. This allows us to formalize the representational capacity an agent can call upon in approaching tasks in terms of information. We find trade-offs occur when components of a task conflict in ways that cannot be easily coarse-grained by the agent's encoding scheme. Further, cognition may be general, specialized, or implement a coverall strategy, depending on conditions.

Human communication is multimodal, characterized by the use of speech and co-speech iconic gestures. While previous research examined how cognitive and communicative demands affect multimodal language use, the interplay between speaker-related cognitive constraints and listener-oriented adaptations remains unclear. The present study examines how speakers with varying spatial skills adjust their speech and iconic gestures when addressing interlocutors with different spatial skills. By employing the imagined addressee paradigm, twenty-three participants described how to solve mental rotation problems to interlocutors with low- versus high-spatial skills. Speakers produced more iconic gestures when addressing low-spatial compared to high-spatial interlocutors. However, this adaptation was primarily observed among individuals with above-average spatial skills. These findings extend prior work on multimodal audience design, showing that while speakers adjust their gestures based on listener characteristics, these adaptations are constrained by the speaker's cognitive capacities. This study highlights the dynamic interaction between cognitive and communicative demands in multimodal language.

Reasoning about the identity of objects is challenging, because objects can experience alterations that change its properties. Researchers have proposed that causal explanations are important in making identity judgments, and one important causal factor is objects' teleology (purpose or function). This study focuses on how teleological information affects identity judgments of entities that split into two descendants. In Experiment 1, we provide evidence that two types of teleological information – function type (structure-dependent or not) and function preservation influence the likelihood that a descendant is judged as the original entity. In Experiment 2, we show that object/substance construal is a mediator of function type in the identity task, suggesting that some of the teleological effects can be explained via construing the entity as object/substance. Together, these two experiments highlight the importance of teleological information in identity judgments.

Spatial language is an integral part of everyday communication, but existing theories fail to explain the cognitive processing underlying the interpretation of spatial descriptions. The traditional spatial template theory does not attempt to provide a mechanism behind the computation of acceptability ratings, nor does it adequately address the effect of distractors or the conversational context. In this study, we propose a new model of spatial pragmatics, based on the Rational Speech Act (RSA) framework. This Bayesian model shows how the patterns of spatial language understanding follow directly from the principles of pragmatics. Our model accounts for previously found effects of angle, distance, and distractors on acceptability ratings and indication behavior. We test these predictions in an experiment consisting of various tasks. The results largely support our model's predictions, suggesting that pragmatic reasoning might play a role in spatial language use.

This study investigates how goals influence information sampling strategies in active learning. Previous work in this area compared different sampling heuristics while holding constant participants' goals (e.g., the final incentivized test). In a behavioral experiment, we examine the effect of generation-driven versus discrimination-driven goals on information-seeking sampling strategies by manipulating the (pre-declared) test condition across subjects. Our results suggest that goals affect information sampling, with discrimination-driven tasks leading to more sampling around class borders (``Label-Margin" sampling strategy) and generation-driven tasks leading to more sampling around class centers (``Most-Certain" sampling strategy). Moreover, we show that strategies evolve over time and are related to the performance of participants. These findings highlight the importance of considering goals in understanding human information-seeking behavior.

People often fall into learning traps where a false belief about the structure of the environment leads to under-exploration of rewarding options. Two studies (N = 324) examined whether observation of the approach decisions of another learner facilitated escape from a trap. After an initial learning phase in a task where approach of different category members could lead to gains or losses, we identified whether participants had learned an optimal two-dimensional categorization rule or fallen into the trap of using a one-dimensional rule. Participants then observed the approach decisions of another learner using the same or a different category rule. Participants' categorization rules in a final round of category learning were then assessed. A substantial proportion of those who had initially fallen into the trap shifted to the optimal rule after observing use of an alternate rule. This effect was found following observation of both optimal (Experiment 1) and sub-optimal rules (Experiment 2). In contrast, those who learned the optimal rule in the initial learning phase were unaffected by social observation. The results show that social learning is a viable approach for facilitating escape from learning traps.

As large language models (LLMs) are increasingly integrated into our lives, concerns have been raised about whether they are biased towards the values of particular cultures. We show that while LLMs were biased toward the values of WEIRD populations, some non-Western populations, including East Asia and Russia, were also represented relatively well. Notably, the Rich dimension was the strongest predictor of LLM's alignment instead of the most discussed Western dimension. This suggests the need to attend to less prosperous populations instead of focusing only on easily accessible populations. We also found that one source of this bias could be unbalanced training data as approximated by an Internet Freedom measure, and that prompting the model to act as individuals from different populations reduced the bias but could not eliminate it. These findings raise the importance of training process disclosure and the consideration of culture-specific models to ensure ethical usage of LLMs.

Future theories of cognition need to encompass a wide range of human experiences, beyond those typically assessed in the laboratory. This study assessed how the experience of a prior artwork (prime) influenced the experience of the next artwork (target), and how this influence was affected by the presence or absence of event boundaries in a VR environment. We found that when primes were more emotionally intense, targets were rated lower in liking/beauty and emotional intensity. This influence was attenuated when the paintings were separated by an event boundary (separate rooms). Surprisingly, there was an effect of event boundaries on the processing of the prime paintings. An evaluation of additional data suggests that this is due to the mere presence of another painting in the same room, even before it is actually viewed. Thus, event structure can meaningfully impact the experience of viewing art.

Verbs show different statistical preferences for syntactic structures. An influential theory for how verb-syntax links are learned suggests that learning is based on, and proportional to, prediction error. However, the evidence is mixed and there is need for evidence from a diverse set of paradigms. We exposed 90 college-aged adults to an artificial language containing novel verbs and sentence structures and tested their production of utterances in the new language. We reversed verb-syntax links from one training block to another and found that participants were able to learn the reversal, as seen in a production test. Pupillometry detected surprise when participants heard sentences in the second training block that had the opposite verb-syntax links than in the first training block. Despite detecting both surprise and successful learning, we did not find an association between the two. Thus, we did not find evidence that learning was based on surprise. We discuss alternative learning mechanisms that can help language users adapt their language based on the input.

Contextual predictability drives both word form and word choice in language use. The effects of the predictability of a word given its previous context are generally well-understood in both production and comprehension, but studies of naturalistic production have also revealed a poorly-understood backward predictability effect of a word given its future context, which may be related to planning. In this study, we revisit backward predictability effects using improved measures and more powerful language models, and introduce a principled measure of planning based on the pointwise mutual information between the word and the future context after controlling for the effects of previous context. We evaluate both measures for predicting word duration, and then extend the scope of these effects to a novel paradigm that involves predicting substitution errors in naturalistic productions. Our findings reveal that the proposed PMI-based measure of planning performs comparably to backward predictability. This analysis provides a useful test-bed for probing the link between past and future context predictability and underlying cognitive processes. Keywords: Language production; Information-theoretic linguistics; Corpus Research;

Both fictive change and motion expressions are linked to dynamic conceptualization, a central concept in cognitive linguistics. However, it remains unclear whether producing these expressions involves the mental simulation of change or a dynamic perception of events—an area that invites further exploration. In this paper, we examine co-speech gestures in a spatial description task, exploring two main predictions: (1) If fictive expressions involve some form of dynamicity or simulation of change or motion, speakers will gesture more frequently than with factive expressions; and (2) If fictive expressions involve dynamicity or simulation, the gestures will reflect this imagery and may be more dynamic than static. The findings from this study support both predictions, suggesting that fictive expressions indeed involve a dynamic conceptualization or simulation of static spatial concepts.

From infancy to adulthood, habituation and dishabituation enable learners to filter out repetitive information and orient to novel information. Because variability in these processes has been linked to differences in later cognitive outcomes, studying individual differences in habituation and dishabituation is crucial for building a more comprehensive model of early learning. Here, we leveraged large-scale datasets spanning infants, preschoolers, and adults to examine how individual differences in habituation predict dishabituation magnitude. We found that faster habituation and higher volatility predicted stronger dishabituation. Moreover, we showed that different measures of dishabituation sometimes yielded divergent patterns, suggesting that measurement choices can influence observed effects and should be carefully considered in developmental research. These findings reveal how endogenous factors are meaningful drivers of looking behaviors. Overall, our results underscore the need for large-scale data approaches to studying visual attention across the lifespan.

Understanding infants' natural input is essential for advancing theories of cognitive development and learning. Recent research indicates that across modalities, infant input approximates a near-Zipfian distribution, with a large amount of input about a few items and substantially less about the rest. However, prior work has only examined aggregated distributions across subjects, focused on a single modality in isolation, and considered the input available to infants rather than what they actively select. We show that at both the corpus and individual levels, infant attention selection and the verbal input infants receive from parents follows a near-Zipfian distribution. Moreover, when integrating across modalities, the verbal input infants hear while attending to the same object becomes even more skewed than verbal input alone. Findings suggest that Zipfian-like structure is not only a property of infant environments but emerges through active selection, highlighting its potential role in shaping early learning.

Exploring the environment efficiently and exploiting learned information effectively are crucial to intelligent agent behavior. Prior work has shown that humans can manage the exploration-exploitation trade-off not only action-by-action, but also at the strategy or rule level, using a heuristic on rule certainty (Collins & Koechlin, 2012). We evaluated this theory on a partially observable rule-switching task and collected human behavioral data (n=112) on two task variants with different levels of rule complexity to test whether taxing cognitive resources impacts exploration heuristics. Our results replicated previous findings, showing that the model is robust to dynamically switching task structure and increased executive demands due to rule complexity. Additionally, we identified a novel meta-heuristic of using high-level rule structure to inform decision-making and computationally characterized its integration with Bayesian inference to support efficient exploration. Through modeling analyses, we show that increased demand on executive function might interfere with this meta-cognitive process.

Text simplification is a common practice for making texts easier to read and easier to understand. To which extent does it achieve these goals, and which participant and text characteristics drive simplification benefits? In this work, we use eye tracking to address these questions for the first time for the population of adult native (L1) English speakers. We find that 42% of the readers exhibit reading facilitation effects, while only 2% improve reading comprehension accuracy. We further observe that reading fluency benefits are larger for slower and less experienced readers, while comprehension benefits are more substantial in lower comprehension readers, but not vice versa. Finally, we find that high-complexity original texts are key for enhancing reading fluency, while large complexity reduction is more pertinent to improving comprehension. Our study highlights the potential of cognitive measures in the evaluation of text simplification and distills empirically driven principles for enhancing simplification effectiveness.

The most commonly used measure of spatial cognition to assess both adults and children is mental rotation. However, little is known about the cognitive strategies that children use to solve this task. Understanding how and when children may employ different mental rotation strategies can illuminate the development of mental rotation ability and help clarify previous mixed findings on the developmental trajectory of children's mental rotation skills. Thus, in this study, we investigated what strategies children use in a new mental rotation task and whether their strategy use would be influenced by the test instructions. In Experiment 1, we found that the types of strategies children used in our mental rotation test differed from strategies reported in previous research, suggesting that strategy use is dependent on test design. In Experiment 2, we found that children's strategy use can be malleable; changing the test instructions reduced one type of erroneous strategy, flipping. Our findings suggest that different tests labeled "mental rotation tests" may actually be measuring different abilities.

The majority of evidence on the relations between young children's levels of food neophobia (the fear of novel food), categorization abilities and executive functions is cross-sectional, leaving the direction of causality unclear. This study aimed to examine the bidirectional relations between children's food neophobia, categorization performance and strategies, and executive functions (working memory, inhibition and cognitive flexibility) longitudinally. Children (n = 113; M age = 48.30 months at Time 1) were assessed at two time points over the course of a year of schooling. Controlling for age, early levels of food neophobia significantly predicted lower subsequent categorization performance and executive functions. No significant evidence was found to support the reverse directionality; neither categorization performance, strategies, nor executive functions at Time 1 predicted subsequent levels of food neophobia. The findings provide longitudinal evidence that neophobia hinders the development of categorization and executive functions abilities.

People's beliefs sometimes diverge after observing the same information, which has been interpreted as evidence of irrationality. This behaviour has been proposed to result from people's limited cognitive resources and motivated reasoning, but how belief revision differs across these explanations has not been formalized or compared to a rational norm. Further, while people may be biased relative to a normative ideal, they may still make optimal choices given their limited cognitive resources, or rationally balance the utility of holding accurate beliefs with the belief's intrinsic utility. Across two studies, we develop and test a unified computational account of belief polarization under these proposed mechanisms, showing that people's performance on a belief updating task best fits a limited-resource Bayesian model; external motivations may contribute to divergence (or convergence) by determining what pre-existing information people consider relevant to a situation, rather than by changing how people evaluate new information in isolation.

Time is often conceptualized spatially. Some argue that this spatialization is essential for understanding time and should manifest reliably in temporal gesture. Here we ask whether people vary in their production of temporal gesture and what this variation might signal about abstract reasoning and communication. Participants (N = 94) watched time-travel narratives, reasoned aloud about the narrative's temporal structure, and later completed a computer assessment of their recollection and comprehension of the narrative's complex temporal structure. We found that participants varied considerably in how often they produced temporal gesture. This variation in temporal gesture was strongly associated with the production of non-temporal representational gesture, suggesting a ‘gestural style' that governs the production of gesture in general. There was no association, however, between temporal gesture and temporal reasoning accuracy. We speculate about the role that temporal gesture might play in larger assemblages of temporal understanding and communication.

To understand how people vary in their cognitive control engagement, researchers use different laboratory tasks and compare performance on trials that are more versus less control-demanding (e.g., congruency effects). However, previous research has struggled to uncover consistent patterns of correlation across cognitive control tasks, leading to questions about the utility of these tasks and the existence of task-general control. The current study sought to test whether these validity concerns may center on the stimulus-driven nature of congruency effects, rather than the tasks themselves. To overcome this obstacle, we varied task incentives while holding stimulus features constant. We show both theoretically and empirically that the effects of incentives on control allocation correlate across tasks. Together, findings support task-general control processes that operate across different contexts.

Overparenting undermines children's self-efficacy and motivation. However, little research has explored whether its negative impacts extend beyond the home and affect not only overparented children, but also their peers. Here, we test the hypothesis that 6- to 8-year-old children attribute peer success to internal (ability) rather than external (parental intervention) causes and that this attribution leads children to form negative beliefs about their own competencies. In Experiment 1, children were more likely to spontaneously attribute outstanding peer performance to internal causes (ability) than external ones (parental intervention). In Experiment 2, children reported lower self-perceived abilities when they learned that peers outperformed them due to internal (ability) versus external (parental intervention) causes. Together, these findings reveal an unintended consequence of overparenting: Intervening to enhance one child's performance leads peers to feel worse about their abilities, potentially harming their self-concept and future motivation.

Humans use clocks to objectively measure their subjective temporal experiences. But can the spatial properties in a clock distort our experience of time ? This study examines how spatial boundaries in an analog clock influence prospective temporal judgements. We found that when the second hand crossed more boundaries, it distorted participants' spatial memories, causing them to overestimate the arc traced by the second hand. However, this distortion in spatial memory did not significantly influence participants' temporal judgments. Besides boundary crossing, our study examined and replicated the influence of speed on temporal judgment, with faster speeds of the second hand being associated with a dilated temporal experience.

Working memory (WM) is essential for planning and decision-making, enabling us to temporarily store and manipulate information about potential future actions and their outcomes. Existing research on WM, however, has primarily considered contexts where stimuli are presented simultaneously and encoded independently. It thus remains unclear how WM dynamically manages information about reward and value during planning, when actions are evaluated sequentially in time and their cumulative values must be integrated to guide choice. To address this gap, we developed an information-theoretic model of WM allocation during planning, implemented using variational recurrent neural networks. In this model, an agent optimizes plan quality while maintaining reward information under WM constraints. To test our model, we designed a task in which participants sequentially observed the rewards available at different future states before executing a sequence of actions, attempting to maximize cumulative rewards. Our results suggest that humans preferentially maintain rewards that are most informative for plan selection, integrating both local and global factors. These findings bridge theories of WM limitations with models of human planning, revealing how cognitive constraints shape decision-making strategies.

Humans rely on social consensus to assess the credibility of information. Confidence in a claim can be influenced by repeated exposure to the same source (dependent consensus), which can create an illusion of consensus. This study investigated whether people differentiate between dependent consensus and consensus derived from multiple independent sources. In two experiments, participants rated their confidence in a claim after exposure to an increasing number of supporting claims within a mock social media environment. Results revealed that although dependent and independent consensus were weighted similarly when exposed to a small number of claims, independent consensus carried greater influence as claim exposures increased, regardless of the claim's valence. These findings were further supported by analysis of free-text justifications of responses using large language models (LLMs). Our findings show that people differentiate between the epistemic weight of different consensus types on social media, where repeated exposure to claims is common.

In the languages of the world, it is more common to form complex words by adding suffixes to the end, rather than prefixes at the beginning. It has been argued that this pattern may reflect the salience of word beginnings (Hawkins and Cutler 1988, Hupp et al. 2009). For example, Hupp et al. (2009) find that English speakers rate sequences of syllables that differ at the end as more similar than those that differ at the beginning. However, subsequent research has shown that people's perceptions of sequence similarity are affected by the word-formation patterns in their native language. While the beginnings of sequences are perceived as more salient by speakers of suffixing languages (e.g., English), the ends are more salient to speakers of prefixing languages (e.g., Kîîtharaka, Martin and Culbertson 2020). Thus, it remains unclear whether universal perceptual preferences are linked to the predominance of suffixing in the world's languages. We address this question by investigating perceptual-similarity judgments in speakers of Cantonese – a language with little affixation. We find that, like English speakers, Cantonese speakers perceive the beginnings as more salient, in sequences of shapes and syllables. This finding revives the possibility of a universal perceptual bias, albeit one that can be strengthened or attenuated with language experience.

Harmonic expectation is an important generator of musical experience often explained through mechanisms of statistical learning. EEG research has identified ERP components associated with expectation, including the Early (Right) Anterior Negativity (E(R)AN), which is theorized to index harmonic surprisal with reference to long-term memory of the statistical structure of music. However, the role of top-down influence remains under-explored. We present data from a novel paradigm that cues listeners to the syntactic structure of the stimuli (but not whether they contain improbable events). Our main result revealed larger E(R)AN amplitudes for surprising chords when listeners knew that additional context would follow a surprising harmony. We propose that listeners prospectively integrate surprising chords with anticipated future context, rather than responding to them solely through automatic probability assessment. Musical surprisal arises from a dynamic interplay between bottom-up cues and a listener's top-down anticipated syntactic structure.

This paper examines how LLMs handle false presuppositions and whether certain linguistic factors influence their responses to falsely presupposed content. Presuppositions subtly introduce information as given, making them highly effective at embedding disputable or false information. This raises concerns about whether LLMs, like humans, may fail to detect and correct misleading assumptions introduced as false presuppositions, even when the stakes of misinformation are high. Using a systematic approach based on linguistic presupposition analysis, we investigate the conditions under which LLMs are more or less sensitive to adopt or reject false presuppositions. Focusing on political contexts, we examine how factors like linguistic construction, political party, and scenario probability impact the recognition of false presuppositions. We conduct experiments with a newly created dataset and examine three LLMs: OpenAI's GPT-4-o, Meta's LLama-3-8B, and MistralAI's Mistral-7B-v03. Our results show that the models struggle to recognize false presuppositions, with performance varying by condition. This study highlights that linguistic presupposition analysis is a valuable tool for uncovering the reinforcement of political misinformation in LLM responses.

In the Gricean tradition, pragmatic competence is part of the general human capacity for social reasoning. Indeed, human performance in reference games involving ad-hoc implicatures sometimes aligns with idealized models of rational interaction. But such experiments have also found that humans derive far fewer implicatures than ideal models, subject to individual differences unrelated to social reasoning. In this paper, we consider whether these patterns could arise from the resource-rational deployment of a core social competence, such that individuals choose from various strategies of interpretation, given those strategies' resource demands and success rates, subject to individually-varying predispositions and exploration tendencies. We construct a model of this resource- rational performance in the cognitive architecture ACT-R—to our knowledge the first mechanistic model of performance in these tasks—and we examine its predictions for multi-trial reference games across two model experiments. The model reproduces the key patterns in the human data, providing an initial proof of concept for the role of resource-rationality in these tasks and opening a new avenue for understanding individual differences in pragmatic reasoning.

How do infants come to represent people's identities? In two experiments (N = 86), we investigated 7- to 10-month-old infants' abilities to individuate (i) one of their own caregivers and an unfamiliar adult and (ii) people who are either socially engaged or disengaged. Although classic research has found that infants at these ages do not individuate objects (e.g., a duck and a ball), the infants in our experiments individuated people, so long as those people were socially engaging. These findings suggest that infants represent people as individual entities before they do objects, which may support the formation of children's relationships or their evaluations of others.

Research on animal cognition is currently struggling with two main problems. Firstly, mainstream views on reasoning in philosophy of mind and epistemology are unable to explain a growing body of empirical results in animal cognition research. The second problem concerns the interpretation of the same empirical results within comparative psychology, namely, whether they are best explained by appeal to (‘lower') associative or (‘higher') cognitive mechanisms, where neither side is providing convincing explanations. This problem is amplified by a crisis of the dichotomy itself between the ‘associative' and the ‘cognitive'. In response to these two problems, I develop an affect-based construal of cognition, which broadens the conception of cognition to also include affective processes. I introduce the concept of affective representations that exhibit both ‘associative' and ‘cognitive' characteristics. This concept allows for new kinds of explanations of comparative research results, and indicates a future direction to explore human and nonhuman cognition.

Denic and Szymanik (2024, henceforth D&S) argued that recursive numeral systems optimize the tradeoff between lexicon size and average morphosyntactic complexity. In support of this claim, they showed that a broad range of attested numeral systems trade off these two quantities nearly as well as the best of a large set of artificial numeral systems. However, D&S' artificial systems were in some respects not entirely comparable to natural ones. Here, we address this issue by creating a grammar framework that can represent both natural and artificial numeral systems, and we derive both natural and artificial numeral systems from this single framework, which ensures the comparability of the two sorts of systems. We test D&S' original claim under these new conditions, and find support for it. We also explore the proposal that numeral systems might optimize the sum of lexicon size and average morphosyntactic complexity under a fixed weighting of the two terms, and the role of the prior distribution over numbers.

What inferences do infants make from people's helping behavior? Two preregistered studies examined whether 14- & 15-month-old infants expect consistent helping behavior across different social contexts, and whether any such expectations are consistent with inferences about relationships or dispositions. Participants saw one individual help a target social partner move a boulder up a hill, and they saw another individual hinder the same target from reaching the top of the hill. We then tested infants' expectations about which individual was more likely to provide help in the future by measuring how long they looked when both the helper and hinderer provided help in a new context. In Exp1, the target social partner in the new helping context was the same character that appeared in the familiarization events. In Exp 2, the target was a novel character. Infants looked longer when the hinderer, rather than the helper, provided help to the original target in the new context (Exp1). However, infants' looking times did not differ between events when the target was novel (Exp2). The looking patterns between the two experiments were significantly different. Thus, infants use the pro- or antisocial nature of an individual's past actions to generate expectations for future behavior, but do not generalize those expectations to new targets. Together, this suggests that infants primarily infer social relationships, rather than dispositions, from others' helping behavior.

This study provides the first evidence that sentence interpretation/inferences that comprehenders draw about situation typicality are moderated by the effects of communication form and sentence polarity. We examined the interpretations of affirmative and negative sentences describing real-world situations of varying typicality. We manipulated whether a sentence is presented as a description with an omniscient narrator (e.g., The house has a bathroom.) vs. as direct speech with the speaker-addressee relationship identified (e.g., "The house has a bathroom," Emma told her partner.). By comparing sentence interpretation across situations varying in (i) typicality (e.g., house has bathroom/garage/ballroom), (ii) communicative form (description vs. direct speech), and (iii) sentence polarity (affirmative vs. negative), we find that presenting information as direct speech encourages pragmatic inferences beyond world knowledge. We also find that negation per se does not trigger theoretically predicted pragmatic inferences, but with a specific communicative act like direct speech it does.

Causal relationships in the real world can have diverse mechanisms with differing statistical signatures. We investigate whether people can distinguish between causes that merely change the timing of events ("when" causes) and those that bring about or prevent those events ("whether" causes). We designed experiments where the rate of an event varies over time due to one such causal influence. Events were shown in real time in Experiment 1 and as a timeline visualization in Experiment 2. Our results suggest that people are capable of identifying "when" and "whether" causes but with a distinctive pattern of confusability: People confuse Generation with Hastening; and Prevention with Delaying. We develop a Causal Abstraction from Summarizing Events (CASE) model, which explains people's judgments as mediated by their rate-change-event detection. We discuss how this line of research can be extended to study human cognition about dynamic causal influences and its relevance to real-life judgment and decision-making.

Children use a variety of cues to decide who they can trust to be a credible source of information. One such cue is accent. Previous research has attributed accent-based preferences to a bias for in-group members. In the present study, we examine another potential contributor to these preferences: processing difficulty. Four- to seven-year-old children completed a selective word-learning task, in which they were presented with pairs of speakers and needed to choose one to learn a new word from. The speakers differed in accent type – native or non-native – and non-native speakers differed in how difficult their speech was to process. Children were more likely to choose to learn from the speaker whose speech was easier to process, and the magnitude of this effect was linearly related to the processing difficulty disparity between the two speakers: the greater the disparity, the stronger the effect. These findings are the first to demonstrate the role of processing difficulty in children's accent-based selective learning.

Recent evidence suggests that delegating tasks to machines can facilitate unethical behavior, but the psychological mechanisms driving this effect are not yet well understood. This study investigates whether two interventions can mitigate cheating in an algorithmic honesty game: transparency (information about which user input causes which algorithm behavior) and framing (natural language cues about the moral valence of behavior). In a 2 x 2 experimental design, we find that transparency does not reduce dishonest behavior, despite participants actively engaging with and understanding the provided information. Conversely, framing — replacing neutral labels like "maximize profit" with ethically charged terms like "maximize cheating" — substantially reduces dishonesty. These findings suggest that curbing misuse of AI requires confronting users with its moral implications, not just explaining the mechanics.

Two learning mechanisms have been suggested to underlie statistical learning: computation of transitional probabilities and chunking. It remains an open question though what determines which mechanism is used. In this study, we examined whether learning mechanisms are exploited differentially depending on the structure of the input to be learned. More specifically, we investigated whether the strength of the relationships between elements in the input structure and the presence of higher-order relationships influence the employment of the mechanisms. Participants were presented with three different input structures. We measured reaction times in a self-paced statistical learning task and created Bayesian models that formalised different learning mechanisms. The results show that the employment of the learning mechanisms indeed depends on the input structure. Further studies will need to examine a more specific mapping between the input structures and the learning mechanisms.

Accurate classification of sleep stages is crucial for sleep quality assessment, health monitoring, and disease prevention. To effectively extract significant waveform features and capture the interactive coupling of features at different layers from single-channel electroencephalogram (EEG) signals, this study proposes the Granularity-Aware Pyramid Network with Graph Aggregation for Sleep Staging (GPNet) model. Specifically, the model first extracts fine-grained time-frequency features from multi-resolution input signals using the feature pyramid. Subsequently, an adaptive deep attention mechanism is incorporated into the layer with the highest depth-wise information to explore the correlations between local and global features. Finally, graph convolution is employed to learn the coupling interactions among high-level features across multiple layers. Comparative experiments conducted on the Sleep-EDF-X datasets demonstrate that GPNet exhibits highly competitive performance compared to other models. Additionally, GPNet predicts post-sleep recognition speed of negative emotions, revealing a negative correlation with REM(%) sleep and suggesting that sleep mitigates negative effects.

In most of the world's languages, complex noun phrases are created by placing adjectives closest to the noun, and demonstratives farthest away, with numerals in the middle (e.g., "this one short paper"). Theoretical linguistics suggests that this tendency may result from a typical hierarchical structure, in which adjectives form an immediate constituent with nouns, numerals combine with that sub-constituent, and demonstratives combine with the resulting unit. Recent experimental studies also support this idea, showing learners prefer orders that follow this hierarchy (e.g., Martin et al. 2020). However, it is unknown whether this same preference is found in children, who have less experience than adults with these structures, and who might be less sensitive to the hierarchical structure of language. Here, we investigate ordering preferences in 5-6-year-old children. Results show that children, like adults, prefer hierarchy-following orders, strengthening the hypothesis that the prevalence of these orders reflects a universal cognitive bias.

What is the developmental trajectory of language-mediated abstraction skills, and to what extent are these skills influenced by semantics? We address these questions asking children to generate semantic relations of categorical inclusion for words varying in concreteness. Results show that abstraction improves over time, independent of age, with both concrete and abstract concepts organized into hierarchical taxonomies. However, abstract concepts allow shorter ladders, making them harder to categorize, especially for younger children. These findings underscore the distinction between concreteness and specificity as separate dimensions of abstract reasoning, and they lend empirical support to theoretical models that treat these facets of abstraction as dissociable.

Semantic feature norms have been foundational in the study of human conceptual knowledge, yet traditional methods face trade-offs between concept/feature coverage and verifiability of quality due to the labor-intensive nature of norming studies. Here, we introduce a novel approach that augments a dataset of human-generated feature norms with responses from large language models (LLMs) while verifying the quality of norms against reliable human judgments. We find that our AI-enhanced feature norm dataset, NOVA: Norms Optimized Via AI, shows much higher feature density and overlap among concepts while outperforming a comparable human-only norm dataset and word-embedding models in predicting people's semantic similarity judgments. Taken together, we demonstrate that human conceptual knowledge is richer than captured in previous norm datasets and show that, with proper validation, LLMs can serve as powerful tools for cognitive science research.

Metaphors are a creative use of language that conveys complex ideas through abstract reasoning and cognitive flexibility. While prior research has demonstrated that music influences creativity, its specific impact on metaphor production remains unexplored. In this study, 90 adults were assigned to one of three groups—silence, happy music, and sad music—and completed a metaphor production task, generating expressions for emotions (e.g., being happy) and actions (e.g., telling a lie). Participants also completed convergent and divergent thinking assessments to account for individual differences in creativity. Results showed that participants who listened to happy music while doing the task were more likely to produce figurative expressions, with convergent creativity positively predicting their production, while divergent creativity had no effect. Moreover, metaphors produced with background music were generally rated as more novel than those produced in silence, with sad music leading to metaphors with a more negative emotional tone. These findings suggest that extrinsic factors, particularly happy music, can enhance our ability to produce metaphors by boosting the cognitive flexibility required for creative thinking.

A central but unresolved aspect of problem-solving in AI is the capability to introduce and use abstractions, something humans excel at. Work in cognitive science has demonstrated that humans tend towards higher levels of abstraction when engaged in collaborative task-oriented communication, enabling gradually shorter and more information-efficient utterances. Several computational methods have attempted to replicate this phenomenon, but all make unrealistic simplifying assumptions about how abstractions are introduced and learned. Our method, Procedural Abstractions for Communicating Efficiently (PACE), overcomes these limitations through a neuro-symbolic approach. On the symbolic side, we draw on work from library learning for proposing abstractions. We combine this with neural methods for communication and reinforcement learning, via a novel use of bandit algorithms for controlling the exploration and exploitation trade-off in introducing new abstractions. PACE exhibits similar tendencies to humans on a collaborative construction task from the cognitive science literature, where one agent (the architect) instructs the other (the builder) to reconstruct a scene of block-buildings. PACE results in the emergence of an efficient language as a by-product of collaborative communication. Beyond providing mechanistic insights into human communication, our work serves as a first step to providing conversational agents with the ability for human-like communicative abstractions.

Do children understand that people vary in how attentive they are and recognize that people prefer attentive social partners? Across six experiments, we showed participants one agent engaging attentively with a child puppet and another who was distracted throughout the interaction. Across four experiments, four and five-year-olds (total N= 132; overall mean: 4.85; range: 4.0-5.9 years) failed to distinguish the agents. Six and seven-year-olds (total N=131; overall mean: 7.01; range: 6.0-7.9 years) succeeded given repeated interactions but not robustly: fewer than half the children consistently chose the attentive agent. By contrast, adults succeeded given a single demonstration. Children's difficulty was not due to task demands; four and five-year-olds readily distinguished agents who did and did not satisfy the puppet's desires. Thus, although children understand attention as a discrete mental state very early in development, and react negatively when adults are not responsive, children may be relatively insensitive to cues to attention as a continuous mental state.

One source of complexity in counterfactual reasoning is the order in which events are presented within the conditional. Counterfactuals with a backwards order of events (aka ‘backtracking' counterfactuals) involve reasoning backward: from the consequent to the antecedent. We extend on prior experimental work (e.g., Rips 2010), and consider the possibilities adults reason over when they backtrack. We find that adults' reasoning strategies tend to be inconsistent when responding to backtracking questions. In scenarios involving a single causal variable, participants do not generally allow for extraneous changes from the actual world. Furthermore, when reasoning forward along a causal chain, participants do not allow for extraneous changes. However, in backtracking scenarios involving multiple causal variables, participants are at chance in choosing worlds with extraneous changes. We provide novel evidence for the changes allowed from the actual world when backtracking, with mixed support for theoretical claims such as Minimal Networks Theory.

When we produce utterances, we aim to express our message in a coherent way. While listeners generally prefer semantically related constituents to be close together in the string ("local"), certain constructions allow long-distance dependencies (LLDs). There is growing evidence that constraints on LDDs involve information structure, but conjunction is recognized to require its own unique constraints. Here we offer 4 experimental studies aimed to illuminate why conjunctions resist LDDs, by investigating a particular case experimentally in detail: English wh-questions that query only the final conjunct in verb phrase conjunction. The first two studies demonstrate that gradient acceptability is predicted by the extent to which a conjunction expresses a single complex event rather than two separate events. Exp. 3 demonstrates a role for the information structure constraint that holds of LDDs generally: more prominent (less backgrounded) conjuncts combine with wh-questions more easily. A final experiment manipulates construal as a single event and the prominence of a conjunct. Results demonstrate an additive causal role for both factors.

Visuomotor adaptation involves explicit strategies and implicit recalibration, but their interaction remains unclear. Strategies can take two forms: algorithmic strategies, involving mental simulation of motor solutions, and retrieval strategies, which rely on previously successful solutions. These strategies arise from distinct neural circuits, which are likely to influence cerebellar-dependent implicit recalibration in different ways. To explore this, we created conditions favoring algorithmic (visuomotor mental rotation) or retrieval strategies by varying training target set size, as retrieval is limited by working memory. We controlled for generalization and intertrial effects, isolating implicit recalibration. Preparation times confirmed distinct strategy adoption. While the magnitudes of implicit recalibration were similar, generalization breadth was narrower with retrieval strategies, suggesting stricter stimulus-response associations. Algorithmic strategies produced broader generalization. These findings confirm that algorithmic and retrieval strategies impact implicit recalibration differently, and demand that future efforts to characterize the pattern of implicit generalization must account for the unique contribution of different forms of explicit strategies.

Human learners' ability to detect rule-governed elements plays an important role in cognitive functions. While extensive research on the acquisition of regularities among adjacent items has provided robust and reliable evidence, learning structured patterns among non-adjacent components-known as non-adjacent dependencies (NADs)-remains far more tentative and only occurs under specific conditions. A past study by Lu and Mintz (2023) found that human learners need more training exposure to detect NADs in visual sequences of object transformations compared to sequences of human actions, but is unclear why. Building on this work, we present a series of three experiments to investigate whether learning NADs from visual dynamic sequences can be enhanced by maintaining the identifiability of the object throughout its transformations. In addition, we explore the effect of providing auditory information-speech or pure tones-along with the visual object transformation sequences. Our findings demonstrate that (a) NAD learning succeeded when speech cues co-occurred and matched with NAD-type frames but failed in the absence of auditory cues (Experiment 1); (b) pure tones presented contingently with the visual sequences also facilitated NAD learning (Experiment 2); and (c) regardless of whether speech or tones were used as additional cues, adult learners were unable to detect NADs when the relationship between specific auditory stimuli and specific visual object transformation sequences was disrupted (Experiment 3).

Social interaction is a complex, multimodal phenomenon with varying timescales and meaning-making structures. Research in this area has progressed along two largely separate paths: qualitative researchers focus on fine-grained analysis, while quantitative researchers computationally identify broader patterns. To bridge this gap and promote cross-disciplinarity, we developed the Dynamic Interaction and Multimodal Signals (DIMS) Dashboard, an open tool for visualizing multimodal data, enabling a qualitative-quantitative synergy in social interaction research. We overview its development, and conduct a proof-of-concept qualitative-quantitative ("quali-quanti") analysis using neural and behavioral time-series data combined with video recording. Our exploratory case study reveals that 80% of segments with sharply increased neural activations in the right temporoparietal juncture align with highly engaged interaction, while 20% correspond to topic transitions. Through triangulation with qualitative insights, we observed that social brain synchrony relates in meaningful moments to head motion synchrony. Finally, we discuss how visualization tools like DIMS enhance multimodal, cross-disciplinary research in social interaction, and tool development.

In two novel verb experiments using the visual fixation paradigm, we investigated how Mandarin-learning toddlers employ distributional cues and semantic cues to categorize novel unaccusative and unergative verbs. In Experiment 1, 31-month-old (but not 19-month-old) participants were found to use the word-order cue to categorize two novel verbs VUA and VUE: after hearing "VUA-le NP" and "NP VUE-le" in the training phase, they categorized VUA as unaccusative and VUE as unergative, showing discrimination in looking times between grammatical trials "NP VUA-le" and ungrammatical trials "VUE-le NP" in the test phase. In Experiment 2, 31-month-olds used the semantic cue of telicity provided via novel events to make categorizations: watching a telic event paired with "VUA-le" and an atelic event paired with "VUE-le" led to differentiation between grammatical trials "VUA-le NP" and ungrammatical trials "VUE-le NP". The findings provide evidence for toddlers' ability to extract information from the input and make generalizations in verb learning.

Responsibility requires causation. But there are different kinds of causes. Some are connected to their effects; others are disconnected. We ask how children's developing ability to distinguish causes relates to their understanding of moral responsibility. We found in Experiment 1 that when Andy hits Suzy with his bike, she falls into a fence and it breaks, 3-year-old children treated ``caused'', ``break'' and ``fault'' as referring to the direct cause, Suzy. By 4, they differentiated causes: Andy ``caused'' the fence to break, it's his ``fault'', but Suzy ``broke'' it. We found in Experiment 2 that when the chain involved disconnection, 3-year-olds focused only on the direct cause. Around 5 they distinguished causes, saying that the disconnecting cause ``caused'' an object to break, it's their ``fault'', but the direct cause ``broke'' it. Our findings relate to the outcome-to-intention shift in moral responsibility and suggest a more fundamental shift in children's understanding of causation.

How do people understand linguistic descriptions of inherently and potentially social events, such as to meet and to dance, and how do these interpretations align or differ across languages? To explore these questions, we developed an empirical database (LISADA), containing ratings for 240 verbs in English and German along two social dimensions: mutuality and jointness. While both languages show an overall positive correlation between these dimensions, hierarchical cluster analyses reveal meaningful within- and cross-linguistic differences. Through an exemplary test case, we demonstrate how these differences can provide insights into the linguistic and conceptual representations of social events, focusing on the role of morphosyntactic marking in event construal.

Advice-taking plays a critical role in collaboration. Yet people tend to under-utilize advice, often to their own detriment. We investigate if people's utilization of advice improves when they know the advisor has access to different information compared to them. We examine how individuals integrate advice in an estimation task, where the advisee and the advisor have access to different perspectives of the same problem. We assess how individuals adjust their estimates when presented with estimates from a human advisor and an AI advisor, and when they are given information about the advisor's perspective. Our findings are consistent with egocentric discounting where individuals exhibit a general bias toward their own information. However, this discounting is lower for AI advisors compared to human advisors in our experiment. Our results also show that the advisor's estimate is taken more into account when the advisor has a more favorable viewpoint -- for both human and AI advisors. This suggests a potential for optimizing advice-taking behavior by enhancing people's understanding of the advisor's viewpoint. This study furthers our understanding of advice-taking dynamics with human and AI advisors and the role of perspective-taking in decision-making processes.

LLM-as-a-judge is considered a potential substitute for human evaluation due to its efficiency and cost-effectiveness. However, recent studies indicate that LLM-as-a-judge exhibits systematic biases when comparing candidate answers, including contextual, verbosity, and positional bias. These biases, as we find, mirror human cognitive biases like the anchoring effect and availability heuristic, where intuitive decisions prioritize superficial features over deeper analysis. Inspired by the Dual Process Theory, we propose that LLM evaluations often resemble system 1 thinking, leading to biased judgments. To address this, we introduce PeBC, a Perturbation-Based Calibration framework that shifts LLM evaluations from system 1 to system 2 reasoning through perturbation injection, bias analysis, and rule calibration. Our experiments on the meta-evaluation benchmarks LLMBar-Natural and LLMBar-Adversarial demonstrate that PeBC successfully mitigates evaluation biases, outperforming existing state-of-the-art (SOTA) methods across various test scenarios and achieving better alignment with human judgments.

Large language models(LLMs) exhibit growing potential as autonomous agents, yet their decision-making capabilities in real-world scenarios remain underexplored, particularly in dynamic scenarios where conditions are constantly changing. Most existing benchmarks mainly focus on static environments, which significantly differ from real-world scenarios. Additionally, existing evaluation frameworks lack fine-grained assessments, providing limited insights during evaluation. To address these, we propose FD-Bench a benchmark for evaluating the decision-making in dynamic scenarios. FD-Bench employs a fire evacuation scenario as a representative dynamic setting and decomposes decision-making into perception, prediction, and action stages, enabling granular evaluation of 8 LLMs and different reasoning frameworks. Our results show that LLMs experience a performance drop of over 50% in dynamic versus static scenarios. Inspired by "chunking" principle in Cognitive Load Theory (CLT), our hierarchical prompting strategy demonstrates improved performance in dynamic decision-making tasks. This work provides insights into LLMs' limitations and pathways toward robust real-world deployment.

This study investigates how individuals balance personal and presentational goals - how they want to be perceived by others - in social interactions where those are conflicting. We develop a computational model that construes presentational goals as minimising the divergence between the perceived and desired belief state of their partner. Based on the divergence between how much a person's partner trusts them versus how much they want to be trusted, we predict complex decision-making patterns that cannot arise from solely focusing on maximising a partner's utility. In accordance with our model, participants tended to forego signalling good intentions and prioritised their own goals when they perceived their partner to trust them. Participants were also less concerned about how they were perceived and acted more often in their own interest when their partner was unlikely to change their mind. We show that people are sensitive to the specific belief state of others and can dynamically adjust their decision strategy to trade off presentational and material gains.

This study explores how readers process different types of pun-based jokes by analyzing their responses to various forms of linguistic ambiguity. Specifically, we examined puns rooted in homonymy, polysemy, and the contrast between idiomatic and literal interpretations of idiomatic expressions. Using EEG, we measured ERPs elicited by the ambiguous elements of these jokes and their punchline. These measurements enabled us to assess how distinct ambiguity types influence the comprehension of punchlines. Furthermore, we compared reader responses to puns against nonsensical sentences and straightforward control sentences. We hypothesize that the differences among joke types will manifest in the relative N400 amplitudes associated with punchline words, providing insights into the neural mechanisms underlying humor comprehension while having more control over joke setups compared to previous EEG studies in this field of research.

Over time we accumulate memories for many related experiences. However, it remains poorly understood how this relatedness, or overlap, among learned information shapes how we remember shared and unique features. The current study investigated this question and further asked whether effects of overlap on memory differ in adolescence compared to adulthood, given evidence that memory specificity continues to be refined beyond childhood. We had adolescents (12-13 years old) and adults learn pairs of objects that overlapped with one another to different degrees and then tested their memory for both overlapping and pair-unique features. Across both age groups, we found that greater overlap boosted memory for the overlapping feature but also led to worse memory for unique features. Further, adolescents were more detrimentally affected by high overlap than adults when recalling specific pairs. Our results suggest there may be a trade-off between memory for shared and unique features of overlapping materials and that adolescents experience a greater cost to this trade-off. More generally, we find that the connections among learned information play an important role in how it is remembered.

Planning in the real world involves navigating vast spaces of possibilities, from finding a route through a city to searching for information online. Yet our understanding of human planning has largely come from studies involving small, simplified environments. To bridge this gap, we explored human planning in the context of the \emph{Wiki Game}, where players start on a random Wikipedia article and are tasked with clicking on hyperlinks to reach a target article with minimal steps. We hypothesized that human planners reduce the computational cost of search by employing heuristic-guided and hierarchical search strategies. Analyzing a dataset of over 75,000 games, we discovered several behavioral signatures of heuristic-guided, hierarchical search. We formalized these insights using computational models, including tree search and hierarchical tree search algorithms. We found that our hierarchical tree search model mimicked these behavioral aspects of human navigation. Moreover, the patterns in human thinking times appeared to resemble patterns in the number of search iterations in the hierarchical tree search model. Collectively, these results suggest that humans use a combination of heuristic-guided search and hierarchical decomposition to efficiently plan in large, complex conceptual spaces.

Prior research suggests that a small vocabulary of meaningful words (a semantic seed) aids distributional learning. In two experiments, we show that adults who are exposed to a complex artificial language are better at inferring the meaning of previously unknown pseudowords when they were taught a semantic seed prior to distributional exposure. We further show that the benefit of a semantic seed is driven primarily by using seed words to discover the relationship between distributional and semantic classes. These results have implications for how syntactic bootstrapping begins.

This study investigates how different modeling approaches affect the measurement of loss aversion, a fundamental concept in psychology and economics. Analyzing 10 datasets comprising over 140,000 trials from 686 participants, we compared four prominent methods: Maximum Likelihood Estimation with Prospect Theory, Bayesian Prospect Theory, Generalized Linear Models, and Drift Diffusion Models. While group-level median loss aversion estimates showed consistency across methods, significant differences emerged at the individual level. The analysis revealed substantial individual-level methodological variability in both the magnitude of loss aversion estimates and participant classification. These findings demonstrate the impact of methodological choices on loss aversion measurement and underscore the need for careful consideration when comparing results across studies using different estimation techniques.

Can we derive computational metrics to quantify visual creativity in drawings across intelligent agents, while accounting for inherent differences in technical skill and style? To answer this, we curate a novel dataset consisting of 1338 drawings by children, adults and AI on a popular creative drawing task. We characterize two aspects of the drawings—(1) style and (2) content. For style, we define measures based on ink density, ink distribution and number of elements. For content, we use expert-annotated categories to study conceptual diversity, and image and text embeddings to compute distance measures. We find significant differences in style and content in the groups—children's drawings had more components, AI drawings had greater ink density, and adult drawings were conceptually diverse. We also highlight a misalignment between creativity judgments obtained through expert and automated ratings. Our work provides a novel framework for studying human and artificial creativity beyond the textual modality.

During communicative interactions, language production and comprehension are bounded by the accumulation of a shared communicative context. Besides lexical pacts and simplification of referential expressions, the creation of a shared context allows for the intelligible use of linguistic signals with novel, interaction-specific meanings. Here, we explore whether context-specific language use leads to mutual adjustment of interlocutors' conceptual representations. We tasked dyads with solving a referential communication game, quantifying dialogue-related adjustments in interlocutors' conceptual structures, and coordination dynamics during the interaction. After engaging in the dialogue, interlocutors judged the same set of referents more similarly than other participants engaged in the same task, but not with each other (pseudo-pairs). Exploratory analyses of the structural complexity of unfolding semantic spaces indicate a stronger alignment between interacting dyads than pseudo-pairs. These findings suggest that human communication is supported by structural coordination of conceptual representations of the communicative referents, over and above signal-level alignment.

Habits -- automatic behavioral patterns formed through repetition -- are essential for daily functioning, but can also lead to inflexible behavior. While crucial for understanding both adaptive and maladaptive decision-making, studying habits' computational and neural mechanisms has been challenging due to limited laboratory experiments demonstrating overtraining-induced inflexibility. We developed a novel task with features designed to encourage participants to engage goal-directed (GD) control between trials (interleaving extensively- and minimally-practiced contexts), then naturally release control within trials (hierarchical multi-step trial structure and opportunities to self-correct). Results showed that overtrained participants displayed stronger biases toward behaviors learned in extensively-practiced contexts, evidenced by higher Habit Index values at early response times. This effect decreased at later response times, suggesting participants could override habitual impulses with GD control. Our computational model, characterizing behavior as a mixture of reinforcement-learned policies, reproduced observed behavioral patterns, suggesting that habits can be viewed as goal-directed deployment of overtrained policies.

As the use of naturalistic speech data of children's language experiences increases, the temporal dynamics of the speech environment becomes a more obvious aspect of speech to investigate. To connect the temporal dynamics in child- centered speech to existing experimental work showing that the temporal presentation of items has measurable effects on learning, it is important to develop measures that quantify temporal patterns in speech. The present work explores one such measure, the burstiness metric, and investigates word burstiness and its relationship with frequency, its behavior across different timescales, and whether it can be used to quantify constructs of massed and spaced orders. Our findings suggest that while related, word burstiness is not an index of frequency, that burstiness varies across both words of different lexical classes and timescales, and that it does not appropriately capture massed and spaced temporal patterns. We discuss implications for how this measure may be used for child- centered audio.

Why and how do we incorporate others' judgments when making an aesthetic evaluation? We investigated this question by studying social transmission of aesthetic evaluations via natural language, which conveys richness that more common (numerical) measures may fail to convey, such as the reasoning behind a judgment. Participants in a large-scale study aesthetically evaluated photographs, either independently or after observing testimony from another person. We found that participants formed more similar evaluations to the testimony they observed (than the asocial control). Furthermore, participants who received the same evaluative testimony wrote evaluations that were more similar to each other in content but not sentiment (relative to a matched asocial cohort). This suggests that social influence on aesthetic evaluations may have a greater informational aspect than previously understood.

Studies of segmental co-occurrence constraints have consistently produced evidence consistent with a cross-linguistic anti-similarity bias for consonants (e.g. Frisch et al. (2004), Pozdniakov & Segerer (2007), Walter (2010), Doucette et al. (2024)). The current study tests whether, in spite of this universal similarity avoidance bias, there are cross-linguistically consistent featural harmony biases the world's lexicons. In particular, we test for the presence of a nasal consonant harmony bias, given the fact that categorical nasal consonant harmony is attested in multiple language families and that nasal harmony is phonetically-motivated. A Bayesian negative binomial model of 91 typologically diverse languages' type frequencies for two-consonant words shows evidence of a weak but reliable cross-linguistic bias in favor of nasal harmony, as well as a comparable bias in favor of voicing harmony. The findings also show patterns consistent with a similarityavoidance bias, most notably a strong cross-linguistic bias against coronal harmony. Taken together, these findings support the notion that similarity-based co-occurrence constraints may be feature dependent in cross-linguistically consistent ways, and more generally that featural dimensions are relevant for understanding the role of segmental redundancy in lexicons.

The origin of abstract concepts remains a central question in cognitive science. Empiricists argue that abstract concepts can be learned through a small set of domain-general cognitive mechanisms, whereas nativists argue that most abstract concepts are not learned at all. To explore these opposing views, we examine the abstract concept of number. Drawing on a recent "emergentist" proposal derived from machine learning models, we tested whether stimuli that align with experiences critical for an empiricist view of number perception result in more accurate and less variable numeric estimations. Five- to nine-year-olds were tested on stimuli that conformed to the natural statistics of objects (i.e., number was correlated with features such as cumulative area and spatial position, and the distribution of stimuli was power-law based) or those that did not comply with these statistics. We find little-to-no evidence for predictions of the emergentist theory when presenting children with natural stimuli.

Psychomotor changes, while crucial indicators of depres- sion, remain underrepresented in clinical observations. We examined the relationship between depression and psychomotor behavior by analyzing head-tracking data related to yaw movements made during exploration of 360° emotional videos, alongside valence and arousal ratings on 9-point Likert scale. Symptoms of depres- sion were recorded using the Patient Health Question- naire (PHQ-9). While subjective ratings for valence and arousal showed no differences across depression groups, the head-tracking data revealed novel results. Individu- als with moderate to severe depression exhibited signif- icantly lower scanning speed and standard deviation in yaw movement compared to minimal to mild depression. Although preliminary, these results underscore the im- portance of psychomotor measures in diagnosis, risk as- sessment, and monitoring in psychiatric care, alongside subjective evaluations.

Many theories conceptualize thinking as search through a space of hypotheses. But what if your initial space is wrong? What cognitive skills support abandoning an ineffective hypothesis space and re-construing a problem with the correct hypothesis space? We examined the development of such abilities in n=172 children ages 3-7 years using Stumper riddles, which challenge respondents to explain seemingly impossible situations. We found evidence that children both learned the relevant hypothesis space for different riddle categories and generalized the cognitive strategy across riddle categories. Although older children showed greater overall accuracy, these effects of learning and meta-learning were found even for the youngest 3-5-year-olds. These results suggest a promising method for probing both flexible hypothesis search and meta-cognitive skills. We discuss ongoing plans to characterize individual differences as a way to uncover the underlying mechanisms of creative problem-solving.

This study explored cross-situational word learning in children with and without autism spectrum disorder (ASD). Children learn words by mapping object names in caregiver utterances to objects in their visual field. We examined the confluence of caregiver object naming and child visual attention in children with and without ASD at play. Head-mounted eye-tracking revealed that children with ASD spent less time attending to named objects than typically developing (TD) children. In both groups, learning input improved as children accrued increased looking time to named objects across multiple naming events. However, for objects with high quantity of naming events, TD children had higher quality learning input than children with ASD. These findings suggest that the input statistics of social interaction are less conducive to word learning in children with ASD. This work has important implications for clinical interventions to scaffold word learning.

Many tasks feel like chores, while others are fun. Why? Here we leverage a popular puzzle game, Sokoban, to explore potential sources of variation in how enjoyable different levels of this game are to solve. In Sokoban, players navigate a grid world, pushing boxes onto goal locations while avoiding getting stuck. We first analyzed natural game play statistics (n = 442 puzzles) and found that some variation in enjoyment ratings could be jointly predicted by surface-level features (e.g., puzzle area) and solution complexity. Next, we measured how much participants reported enjoying a puzzle immediately after attempting it (N= 250 participants). We found that on successful attempts, participants enjoyed it more when they took fewer moves, whereas when unsuccessful, having made more moves was associated with greater enjoyment. Together, these studies advance understanding of how both features of the task environment and the dynamics of exploration make some activities more fun than others.

There is a longstanding debate over how much idiosyncratic, item-specific knowledge is contained in our mental grammars, in addition to productive knowledge of item-general rules and constraints. A key source of evidence is ordering preferences for syntactic alternations like the dative ("throw him an apple" vs. "throw an apple to him") vary depending on which words they contain. But the quantitative extent of this variability is poorly understood, especially in relation to superficially similar, non-dative constructions which are not alternating ("throw the man to the floor" vs. "*throw the floor the man"). To address this, we built a large corpus of naturally-occurring sentences including either dative or superficially similar non-dative structures, and analyzed the unique contributions of productive and verb-specific knowledge in predicting argument ordering preferences

How are comprehenders able to extract meaning from utterances in the presence of production errors? The noisy-channel theory provides an account grounded in Bayesian inference: comprehenders may interpret utterances non-literally in favor of an alternative with higher prior probability that is close under some error model. However, we lack implemented computational models of prior expectation and error likelihood capable of predicting human processing of arbitrary utterances. Here, we model sentence processing for ``noisy" utterances as incremental and approximate probabilistic inference over intended sentences and production errors. We demonstrate that the model reproduces patterns in human behavior for anomalous sentences in three separate case studies from the noisy-channel literature. Our results offer a step towards an algorithmic account of inference during real-world language comprehension. Our model code, implemented in Gen, is available at https://github.com/thomashikaru/noisy_channel_model.

Hippocampal subregions and medial prefrontal cortex may differentially shape memory precision in the mature brain: While anterior and posterior hippocampus may store the general themes and specific details of an episode, respectively, medial prefrontal cortex may instead connect related memories. However, given continued change to the functionality of these regions beyond childhood, it is unclear how children's memory precision is influenced by these same mechanisms. We characterized how hippocampal subregions and medial prefrontal cortex separately and in tandem encourage memory precision in childhood versus adulthood. Children (7-9 years old) and adults studied scene photographs and then performed a recognition test that included both the studied scenes and highly similar lures. Behaviourally, adults had more precise memories than children in that they were better able to discriminate studied scenes from lures. At the neural level, anterior hippocampus and medial prefrontal cortex were differently engaged during this memory formation: Anterior hippocampus engagement was related to subsequent memory across age groups, while children showed greater medial prefrontal cortex engagement than adults overall when studying scenes. Considering individual differences in engagement revealed further developmental differences. Children showed evidence for a trade-off in their reliance on posterior hippocampus versus medial prefrontal cortex during precise memory formation, suggesting competition between these regions. By contrast, the same structures in adults played a more cooperative role in supporting memory precision. These findings suggest that the relationship between posterior hippocampus and medial prefrontal cortex reverses over development to yield adult-like memory precision, such that these regions work in opposition in childhood before becoming specialized to cooperatively encourage precision in adulthood.

Widespread cognitive decline in older adults has been hypothesized to stem from a fundamental deficit in inhibition, or the ability to ignore goal-irrelevant information. The extent to which inhibition operates across different timescales, however, has been under-explored. We introduce a novel cognitive task designed to assess both between- and within-trial inhibition using a common set of stimuli. Behavioral results from younger, middle-aged, and older adults (N=100; age range: 18-73) reveal significant age-related differences in between-trial inhibition, with older adults showing less efficient adaptation to rule changes compared to younger adults. Within-trial inhibition, requiring suppression of distractors within the current visual environment, appears to remain intact alongside normal aging. These findings will support the development of tools for the early detection of age-related cognitive decline, prior to subjective awareness of impaired daily functioning.

Remaining oriented while navigating is a key aspect of survival for many mobile organisms. Previous work suggested that the parietal lobes play a key role in helping navigators determine directions to landmarks relative to themselves. Recent evidence suggests that alpha-band oscillations are crucial for spatial attention and may track egocentric direction as well. We used mobile EEG to integrate these disparate lines of research and test our novel "alpha window hypothesis" that alpha-band oscillations support the retrieval of egocentric directional information around navigators oriented within a real-world environment. Time-frequency-based machine learning analysis revealed significant classification accuracy of the target's egocentric direction within the 8-12 Hz frequency range, thus supporting the alpha window hypothesis. Our results provide a pivotal advancement in our understanding of the neural mechanisms of directional memory by extending previous research that used neuropsychology, fMRI, and EEG into the domain of a dynamic, situated, embodied spatial memory task.

Context valence has been shown to predict word learning in adult experiments. Little is known about whether this extends to children. To address this gap, we conducted a pre-registered word learning experiment to investigate how emotional narrative context shapes children's learning of novel adjectives during naturalistic reading. 120 children aged 7 to 11 years from UK primary schools read 15 novel words (such as "garive") embedded in 30 short narratives of either neutral, negative, or positive valence. Three immediate post-tests assessed learning. We found that children were able to learn novel adjectives from reading short narratives, and older children outperformed younger children. Novel adjectives read in more emotional (positive or negative) contexts were recognized more accurately than those read in neutral narratives. The findings extend previous research conducted using noun concepts and with adults, providing further evidence for affective embodiment in supporting the learning of abstract concepts.

Adults often exhibit various emotional responses to children's performance. How do these responses shape children's learning and motivation? This study investigated how overly positive emotional reactions influence children's task engagement and their willingness to take on challenges. Children aged 5 to 7 (N = 81) completed some simple tasks and received either overly positive or mildly positive emotional responses from an adult. Girls were more likely to continue engaging in the task and take on challenges after receiving a mildly positive emotional response compared to an overly positive one. In contrast, boys exhibited the opposite pattern, engaging in the task for longer and taking on more challenges after receiving an overly positive response than a mildly positive one. These findings suggest that girls and boys may interpret varying levels of positive emotional reactions differently, highlighting the complex, gender-specific ways in which emotional cues influence children's motivation and learning.

Human languages vary widely in the combination and balance of phonetic cues used to encode speech sounds, a source of non-native accents in second language learners. It has been hypothesized that repeated exposure to an accent leads to adaptive perceptual changes in native listeners through multidimensional distributional learning. Although this hypothesis is highly influential, it has rarely been tested against the complexity of naturally produced accented speech, and for a period longer than a single-session experiment. The current large-scale (N = 338), five-session experiment goes beyond this status-quo to examine the adaptation of native American English listeners to naturally produced Mandarin-accented English. A repeated exposure-test design was used to characterize adaptive changes in perception from the first few minutes to over the course of three weeks. The results reveal that behavioral changes can be predicted by listeners' sensitivity to changes in phonetic cue distributions. Possible joint contributions of early-stage auditory normalization and later-stage decision processes are discussed.

To resolve moral dilemmas, people often rely on decision strategies such as cost-benefit reasoning (CBR) or following moral rules. Previous studies show that people learn to increasingly rely on whichever strategy led to better outcomes in the past. Do they learn this by constructing a mental model of what outcomes would result from using either strategy (i.e., model-based learning) or by assigning value directly to each strategy (i.e., model-free learning)? To answer this question, we adapted the two-step task to a trolley-type dilemma between following moral rules (e.g., obeying authority) versus CBR (e.g., saving a larger group). In each of the 125 trials, participants' choices led to either a common or a rare transition, which probabilistically led to good versus bad outcomes. Computational modeling and pre-registered analysis of behavioral data provided converging evidence that participants apply both model-based and model-free learning.

Our expectations about what normally occurs influence our explanations. In conjunctive causal structures, people tend to select the more abnormal cause. This tendency reverses in disjunctive structures, and people select the normal cause. It is currently unknown how these tendencies develop, and what factors contribute to their emergence in childhood. Across three experiments, we tested adults (n = 179) and 5- to 7-year-olds (n = 96) on two tasks where an abnormal and normal factor jointly caused an outcome. Experiments 1 and 2 revealed that while adults' explanations varied according to the causal structure, children exclusively chose the normal cause, regardless of causal structure. Using a task with an intuitive and explicit causal mechanism, Experiment 3 found that children were more likely to select the abnormal cause in the conjunctive case than in previous experiments. This suggests that intuitions about causal mechanisms may facilitate adult-like judgments. We consider potential explanations, including the role of counterfactual reasoning.

Recent studies highlight the context sensitivity of image captioning, where the context in which an image appears strongly influences its caption's informativeness and linguistic style. While AI-generated text increasingly resembles human language, its informativeness and coherence, derived from cross-modal image-text reasoning, may still fall short of content generated by human experts. Given the intertwined nature of informativeness and linguistic style, this study examines news image captioning, a naturally high-context task, to manipulate caption informativeness and assess human sensitivity to such variations. Two experiments (N = 378) and logistic regression analyses reveal that while humans effectively interpret informational cues, their intuition about AI linguistic style often diverges from actual AI language markers. Moreover, humans more readily integrate multiple modalities in preference tasks but rely heavily on linguistic-based strategies for AI detection. These findings underscore the adaptability of human evaluation in image-text systems and suggest informative signals as the more reliable basis for judgment.

Humans make decisions collectively by combining individual and social learning. Individuals benefit from groups when individual exploration fails to accurately assess the environment, a phenomenon known as the "wisdom of crowds." Previous studies indicate that self-organizing group dynamics can reduce suboptimal biases in noisy environments, particularly in fully connected groups. However, agents often have only partial information due to cognitive and physical constraints. To explore how diverse social network structures influence the collective dynamics of social learners, we integrate a decentralized network with a social reinforcement learning model in repeated two-armed bandit tasks. Our results suggest that: 1. Social learning in a sparse network outperforms asocial solo learning in highly uncertain tasks. 2. The phenomenon "less is more, and more is different" holds true only when agents strategically balance individual and social learning. 3. The group size effect on collective performance is significantly influenced by network structures.

Spatial memory is a core cognitive function of many mobile animals. The study of spatial cognition is highly interdisciplinary and different approaches have led to discordant hypotheses about spatial memory. The neuroscientific discovery of place cells and grid cells in the rodent hippocampus and entorhinal cortex led to the highly influential "cognitive map" hypothesis. Conversely, behavioral evidence led to the "cognitive graph" hypothesis that suggests memory is distortion-prone. We argue that between-species differences in sensory and perceptual systems cause vision to play a predominant role in primates. We build on previous modeling work by developing a visual version of the SR model to show that it can provide a unified framework to account for seemingly disparate findings from the brain and behavior, thus providing evidence for our hypothesis that primate (e.g., human) spatial cognition is driven by the hippocampal system, which instantiates a visually-guided cognitive graph.

Children are active, curious learners. How might children's curiosity shape their curriculum during word learning? Past research suggests that children's tendency to explore can lead them to discover novel information during learning. This exploratory tendency could be especially useful when learning word meanings: exploring potential meanings for words broadly could help children efficiently probe a word's possible extension. To investigate this question, we tested how children (5-8 years of age) and adults sample information when presented with a novel word and tasked with uncovering the word's extension. Overall, we found that children explored novel word extensions conservatively. Children (as well as adults) favored sampling choices that confirmed a novel word meaning, as opposed to exploring broader possible meanings. Younger children's sampling choices were especially conservative, with children often sampling the narrowest possible generalization option. Older children were more exploratory, probing broader possible word extensions more frequently. Counter to proposals that children are generally more exploratory at younger ages, our results suggest that when children test the extension of novel word meanings, they are often more likely to confirm their hypotheses than to explore.

What determines which stories (or parts of stories) about the social world are captured and conveyed by children? How do they transform with retelling? We use an iterated learning paradigm to explore how peer-to-peer transmission of explanatory stories (here, explanations for the social customs of novel social groups) is influenced by explanatory framework (natural, supernatural, or hybrid) and children's existing belief systems. Our participants were 69 Hindu and Muslim 3rd-7th-graders in Gujarat, India. Consistent with the `minimally counterintuitive' nature of many highly culturally preserved concepts, hybrid explanations (containing both natural and supernatural elements) were transmitted with the greatest fidelity across chains. Individual religiosity also affected transmission: children who reported themselves as more religious transmitted scientific explanations less faithfully (and hybrid explanations more faithfully) than less religious children.

The network science perspective on vocabulary development emphasizes the structured relationships between words and how they can influence learning. Words that appear in many contexts and thus develop associations with many other words tend to be learned earlier---a growth model called preferential acquisition. Likewise, children appear to have a bias towards learning new words associated with many known words---a growth model called lure of the associates. Although both are statistically related to age of acquisition estimates, much variance remains unexplained, and it is unknown what structures these models promote within the developing vocabulary. We simulated vocabulary growth strictly adhering to preferential acquisition and lure of the associates and found that they promote similar structures: more connectivity, more clustering, and much shorter path lengths than random growth would achieve. They clearly promote small-world structure, consistent with that seen in young children's vocabularies.

For decades, social psychologists have wondered about the cognitive foundations of social stereotype use. Arguments have generally centred either resource constraints, framing stereotypes as `energy-saving devices', or `fit', framing stereotypes as tools to represent real structure in the social environment that sometimes go awry. These resource-based and fit-based accounts have typically been presented as being in opposition to one another. In this paper, we seek to show that both are compatible under an information bottleneck model of agent representation. Through a simple simulation experiment, we demonstrate how stereotype use emerges in resource-rational representations as a function of both capacity constraints and the structure of the social environment. We then use the same framework to consider a possible explanation for the outgroup homogeneity bias in terms of limited cognitive capacity.

People's judgments about what an agent can do are shaped by various constraints, including probability, morality, and normality. However, little is known about how these representations of possible actions—what we call modal space representations—are influenced by an agent's knowledge of their environment. Across two studies, we investigated whether epistemic constraints systematically shift modal space representations and whether these shifts affect high-level force judgments. Study 1 replicated prior findings that the first actions that come to mind are perceived as the most probable, moral, and normal, and demonstrated that these constraints apply regardless of an agent's epistemic state. Study 2 showed that limiting an agent's knowledge changes which actions people perceive to be available for the agent, which in turn affects whether people judged an agent as being "forced" to take a particular action. These findings highlight the role of Theory of Mind in modal cognition, revealing how epistemic constraints shape perceptions of possibilities.

This study investigates how adults in the third (60–79 years) and fourth age (80+ years) integrate words across sentence boundaries during reading. Participants read two-sentence passages involving direct lexical repetition or bridging inferences. Both age groups exhibited longer reading times when bridging was required, showing that inference-making is still present despite potential cognitive declines. However, while third-age adults showed immediate sensitivity to inference demands, fourth-age participants demonstrated a delayed response, suggesting compensatory strategies. These findings highlight a key role for semantic knowledge in sustaining reading comprehension in old age. Future research with more diverse samples and longitudinal methods should clarify how age-related changes interact with linguistic resources. Interventions may target processing speeds to support reading comprehension in late adulthood.

From early in life, humans expect members of the same social group to act like one another. What drives this expectation? Across two experiments, we investigated how 8- and 9-month-old infants' (N = 100) expectations about shared behaviors align with accounts based on collective identities, ritualistic actions, or conventions. In Experiment 1, infants inferred that an action would generalize to a new group member only when they had previously seen more than one group member share that action, suggesting that multimember demonstration influences infants' inductive reasoning. In Experiment 2, infants did generalize an action after observing it from just one group member, but only if they had observed that same action shared by two members of another group in a different social context. Together, these findings suggest that infants learn to recognize which actions are socially conventional and then readily generalize these actions even in new social contexts.

While temporal contiguity and order effects have been shown to be highly robust in the word list memory literature, little is known about how the presence of boundaries signaling higher-order linguistic units affects temporal order memory for items within those units. Here, we present the results of two sentence memory experiments that show that linguistic boundaries block contiguity effects through a temporal context mechanism, whereby words encoded across distinct linguistic environments are also encoded as contextually dissimilar. Two consequences of this encoding mechanism are that (i) retrieval of an item results in reactivation of that item's linguistic context alone due to co-activation of contextually similar content and (ii) significant linguistic boundaries reduce encoding interference between proximal, semantically similar items. We take these results to suggest that linguistic groupings map onto encoding contexts, which constrain the effect of item-to-item associations in sentence memory.

A core aspect of language comprehension is predictive processing, which supports real-time inferences under uncertainty. Real-time prediction is constrained by the mind's limited working memory (WM) resources, which are required for maintaining the context that supports prediction, and for reallocating degrees of belief across inferred interpretations as new pieces of information (e.g., words) are perceived. What is the nature of this WM resource? Is it language-specific or shared between language and other cognitive domains? Do both domain-general and domain-specific resources support prediction? Here, we study this question using an individual differences approach. We collected self-paced reading times of naturalistic paragraphs in English and measured, for each participant, their domain-general WM (backwards digit span) and linguistic WM (reading span). We quantified predictive processing as the relationship between surprisal and reading time. We found that surprisal influenced reading times more strongly in participants with (1) stronger domain-general WM, but (2) weaker linguistic WM, although the latter relationship was less reliable. Our results indicate that domain-general WM could contribute to predictive processing during language comprehension. We discuss several theoretical interpretations of this finding, we well as potential reasons for the discrepancy between our results and past studies.

Aging greatly affects memory, but not all aspects are impaired equally. Past research has demonstrated that older adults show greater deficits in remembering where an object was encountered compared to what was experienced. A third dimension, that has received less attention in this context is memory for when events occurred. In this study, we employed a Sequential Memory Task (SMT) in which participants memorized spatio-temporal visual object sequences over repeated exposures. In two experiments with 39 younger (YAs, 18-35 years) and 53 older adults (OAs, 65-75 years), we examined age-related differences in sequence memory (when) and the interplay between item (what), and location (where) knowledge. Our results revealed that memory was stronger for item sequences than for location sequences or item-location combinations in both younger and older adults. In addition, older adults exhibited greater age-related deficits in location-related memory. We also found that item and location memory were bi-directionally related and that even pure item sequence reports involved location memory and vice versa. Both age groups relied more on item-location binding associations than on transition learning, but computational modeling indicated a higher reliance on independent location transition learning in OAs than YAs. This suggests that the strong age-related impairment in spatial location memory was in part driven by age-differences in memory binding. These findings provide insights into age-related changes in spatio-temporal sequence memory, and highlight distinct learning strategies in younger and older adults.

Fuzzy Trace Theory and the Activation Monitoring Framework disagree on whether gist or backwards associative strength best describes false memories in DRM. Recent evidence suggests situational features best describe processes underlying DRM results. Thematic, goal-derived, and relational categories capture different aspects of situational features, but perform differently in memory and coherence. We constructed novel category-specific lists for use in the DRM paradigm to determine whether different aspects of situational features make different contributions to a successful DRM result, whether relational and goal-derived content can produce false recognition despite low gist and backwards associative strength, and whether DRM captures an aspect of category coherence that has been difficult to measure in relational and goal-derived content. Only relational and thematic content produced sufficient false recognition. This provides mixed support for situational features, evidence that relational content makes specific contributions to DRM success, and evidence of coherence in relational categories.

Some impossible things are more impossible than others. Magically levitating a feather seems easier than levitating a rock, even though both are impossible in the real world. But within the things that are inconceivable---e.g., "the number 13 writing a play" or "a girl being a prime number"---are some things more inconceivable than others? We first established that people have graded, systematic judgements of the likelihood of inconceivable and nonsense sentences (Experiment 1). We then examined two hypotheses as to how people make such judgments: the ease of a metaphorical interpretation (Experiment 2), and how difficult it is to transform a nonsense statement into a sensible one, as measured by distance in a type hierarchy (Experiment 3). We found that graded judgments of inconceivability are not captured by metaphorizability, but do correspond to a measure of distance in a type hierarchy. Our results suggest that inconceivability is graded, and the perceived likelihood of an inconceivable event may be a product of one's ontology of the world.

How does power shape communicative freedom? During communication, speakers often balance multiple competing goals—such as being informative versus preserving their reputation. Across four experiments (N=~1400), we examined how the power relation between speakers and listeners affects narrative production in the face of conflicting informational and reputational goals. Participants imagined having to confess about a minor wrongdoing and were assigned to low, equal, or high power roles (an employee speaking to a supervisor, co-worker, or intern). Low power speakers were more likely to prioritize informativeness and believability. Importantly, the open-ended narratives they provided were more homogenous than those of high or equal power narrators, suggesting that power restricts the range of utterance choices considered. Speakers' ratings of pre-written narratives indicated that power may additionally affect how speakers evaluate the believability of different utterance choices.

Rejection from clubs, teams, or schools is an inevitable part of growing up. Knowing when to quit or persist in the face of rejection is critical for goal pursuit, yet it is unclear how children respond to various sources of rejection. In a pre-registered experiment (N = 202), we tested whether 7- and 8-year-old children are sensitive to the cause of rejection. Children played a game in order to try out for a selective club and were rejected either based on merit (their performance) or by chance (a spinner). Children who experienced luck-based rejection felt better about their competence and persisted marginally more than those rejected based on merit. Across conditions, girls persisted more than boys, and persistence declined with age. These results suggest that by early elementary school, children are sensitive to the cause of their rejection, with implications for how they calibrate effort and pursue goals.

When addressing infants, parents often exaggerate positive affect in both their faces and speech. While these cues have been theorized to support learning, their frequency alongside linguistic input and the information they convey in children's real-world experiences remain unclear. We analyzed ~377 hours of egocentric at-home videos from infants (5–28 months) to examine affective cues in faces and language surrounding early-learned words. Using automated tools, we tagged happy affect in the utterance in which each word was embedded and in co-occurring faces. Faces were visible in only 13.5% of word instances, and even fewer displayed happy affect. However, more (vs. less) positive words tended to co-occur with happier faces. Linguistic context, in contrast, conveyed stronger positive affect and more reliably aligned with word valence than facial affect. These findings suggest that facial and linguistic affect may serve distinct roles in infants' learning environments.

It has recently been theorized that languages evolve under pressure to attain near-optimal lossy compression of meanings into words. While this theory has been supported by broad cross-linguistic empirical evidence, it remains largely unknown what cognitive mechanisms may drive the cultural evolution of language toward near-optimal semantic systems. Here, we address this open question by studying language evolution in the lab via iterated learning. Across two qualitatively different domains (colors and Shepard circles), we find that semantic systems evolve toward the theoretical limit of efficient lossy compression, and over time, converge to highly efficient systems. This provides direct evidence that adult learners may operate under a bias to maintain efficiently compressed semantic representations. Moreover, it demonstrates how this bias can be amplified by cultural transmission, leading to the evolution of information-theoretically optimal semantic systems.

In many languages, words in count syntax quantify over countable individuals (e.g., too many strings), while mass nouns often don't (e.g., too much string). Theories differ in how to characterize nouns that violate this pattern, such as object-mass nouns (e.g. furniture, clothing). These nouns exhibit mass syntax, but often quantify by number (Barner & Snedeker, 2005). On one hypothesis, the individuation of object-mass nouns is lexically specified (Bale & Barner, 2009). Another argues that, while count nouns always quantify by number, object-mass nouns have different quantification criteria depending on context (Rothstein, 2010), including function fulfillment (McCawley, 1975). We evaluated these hypotheses by comparing English quantity judgments for object-mass nouns to (1) superordinate count nouns, and (2) French judgments for translations of object-mass nouns. In each case, we found that object-mass nouns behaved like count nouns, and were no more susceptible to contextual effects. These findings support the idea that object-mass nouns specify individuation lexically.

The think-aloud method, where participants voice their thoughts as they solve a task, is a valuable source of rich data about human reasoning processes. Yet, it has declined in popularity in contemporary cognitive science, largely because labor-intensive transcription and annotation preclude large sample sizes. Here, we develop methods to automate the transcription and annotation of verbal reports of reasoning using natural language processing tools, allowing for large-scale analysis of think-aloud data. In our study, 640 participants thought aloud while playing the Game of 24, a mathematical reasoning task. We automatically transcribed the recordings and coded the transcripts as search graphs, finding moderate inter-rater reliability with humans. We analyze these graphs and characterize consistency and variation in human reasoning traces. Our work demonstrates the value of think-aloud data at scale and serves as a proof of concept for the automated analysis of verbal reports.

Humans are adept at setting goals quickly and flexibly in their daily lives. Previous research has shown that people can assign rewarding properties to abstract or novel outcomes and use them to guide behavior. However, the mechanisms supporting this flexibility and their impact on learning processes, such as working memory (WM) or slower incremental systems, remain unclear. To address this, we designed an instrumental learning task in which participants learned stimulus-action associations by pursuing either standard goals (+1) or counterintuitive goals (+0) under varying WM loads. Our behavioral and modeling results revealed that when pursuing counterintuitive goals, humans learned more slowly and shifted their reliance from WM to habit-like associative processes, despite both processes remaining functionally intact. Additionally, we replicated previous findings showing that humans do not rely on reinforcement learning (RL) processes but instead integrate WM and habit-like processes to learn the associations. This interplay between WM and habit-like processes may allow a more resource-efficient approach to pursuing diverse goals. Our findings shed light on the breadth and cost of people's ability to flexibly learn and pursue any goal.

Do children see emotional consolation during times of hardship as a cue for close relationships? In this study, 6- to 8-year-old children in the U.S. (N = 62) were presented with vignettes in which a protagonist experiences a hardship. The protagonist then tells one of two side characters (i.e., their 'friend') that they either felt sad or okay about the situation, and the character either hugs or does not hug the protagonist. Children inferred that side characters who consoled when the protagonist felt sad were (i) better friends with the protagonist, (ii) more likely to be shared the protagonist's secret, and (iii) more likely to be reciprocated emotional support by the protagonist in the future. Together, these findings suggest that children see consolation during hardship as a cue for social affiliation and may use emotional support to differentiate among positive social relationships.

Lexical ambiguity has classically been categorized into two kinds. Homonyms are single word forms that map to multiple, unrelated meanings (e.g., "bat" meaning baseball equipment or a flying mammal). Polysemes are single word forms that map to multiple, related senses (e.g., "breakfast" meaning a plate of food or an event). Yet there is a longstanding debate as to whether polysemy and homonymy reflect distinct cognitive representations. Some (e.g., Fodor & Lepore, 2002; Klein & Murphy, 2001) posit that they do not—merely describing differing patterns of usage—while others (e.g., Falkum & Vicente, 2015; Pietroski, 2018) argue that polysemes, but not homonyms, involve an underspecified representation that is neutral with respect to the form's multiple senses. While some extant experimental evidence supports the latter view (Klepousniotou, Titone, & Romero, 2008; Srinivasan, Berner, & Rabagliati, 2019), there has not yet been clear evidence of the representation of lexical ambiguity affecting domain-general reasoning. Using a novel inference paradigm, we compare participants' dispositions to endorse deductive, Aristotelian arguments with equivocating polysemes versus comparable arguments with equivocating homonyms. We find that participants endorse the former substantially more than the latter, a phenomenon that we dub the Uncommon Sense Effect. Our results provide direct evidence that polysemes and homonyms have underlyingly distinct mental representations—in particular that polysemes uniquely invoke an underspecified representation

We explore the impact of crosslinguistic influence in first language (L1) attrition, changes in an individual's L1 due to exposure to additional languages. We report an experiment examining reference production in Mandarin in a picture description task by native Chinese speakers residing in Italy or Spain. Mandarin allows null subjects, where subjects can be expressed with a null or overt pronoun; previous work shows that L1 Mandarin speakers exposed to English use more overt pronouns in Mandarin than their more-monolingual peers. In the study reported here, despite exposure to two languages (Italian and Spanish) that, unlike English, allow null subjects, our multilingual speakers used fewer null pronouns and more overt pronouns than their more-monolingual Chinese peers. These findings contribute to attrition research by disentangling the impact of crosslinguistic influence in L1 attrition, and provide insights into the effect of bi- and multilingualism on linguistic systems.

Children are keen explorers of the outside world: They systematically explore surprising findings and test hypotheses during play. However, less is known about whether toddlers are similarly driven to explore and learn about the self. The present work adapts classic exploratory play paradigms to ask whether toddlers are intrinsically motivated to explore their own competence. In Experiment 1, we selected Montessori practical life toys that were verified to be developmentally appropriate and equally appealing to toddlers (N = 24, ages 24-35 months). In Experiment 2, 2-year-olds (N = 49, ages 24-35 months) played with these toys along with a parent. Toys were presented in pairs. In each pair, the parent guided the toddler's hand while playing with one toy, which provided confounded evidence about the toddler's competence, and took turns playing the other toy independently, which provided unconfounded evidence. When given a chance to freely explore the toys on their own, toddlers first approached the confounded toy, which suggests that toddlers sought to resolve uncertainty about their competence. As a further test of this idea, Experiment 3 (N = 11, ongoing) asks whether toddlers' exploration is modulated by task difficulty. Preliminary results suggest that toddlers explore confounded toys more often for more challenging tasks compared to easier ones. Together, our work provides insights into children's early motivation to understand the self, and this understanding is an important first step for researchers, educators, and parents to better encourage and scaffold this motivation throughout development.

Artificial intelligence (AI) can enhance human decision-making by providing assistance at different levels of abstraction. This study investigates whether AI should offer broad, high-level guidance (global AI) or focused, low-level assistance (local AI) to optimise performance and learning. Using a hierarchical multi-armed bandit task where both AI types provide equally valuable recommendations, we evaluate how participants leverage AI support in making sequential decisions. Findings reveal that while participants benefited from both types of AI suggestions, global AI led to significantly greater performance improvements. These results contribute to our understanding of human-AI interaction in hierarchical problem-solving, highlighting the importance of designing AI systems that effectively support human cognitive processes.

Why are nouns often over-represented in children's early words? Prior research has suggested a range of possible explanations for the observed bias in the syntactic category composition of early vocabularies, including cognitive, linguistic, and contextual factors. However, these factors are not mutually exclusive and can be difficult to disambiguate empirically. Across three analyses of the vocabularies of children acquiring two languages (N = 1997), we investigated the role of different language combinations, different levels of language exposure, and different environmental contexts, finding some support for cross-linguistic modulation of syntactic category bias, as well as variation across countries. These results suggest that both linguistic and contextual factors contribute to syntactic category biases in young children's vocabularies, and highlight the importance of understanding the broader background of the child when studying language acquisition.

The human brain represents structure in the external world as "cognitive maps". However, it remains unknown how it represents structure in one's own behaviour. Recent findings show rodent medial-prefrontal cortex (mPFC) does this with a structure-sensitive representation, where all future actions are represented simultaneously. Here, using 7T fMRI, we test for this representation and its properties in humans. Using RSA and a computational model, we show this representation in mPFC and orbitofrontal cortex, while entorhinal and orbitofrontal cortices contain a pure abstraction of task structure. Preceding the future actions representation, action plans are ‘loaded' to mPFC once subjects were given all action-relevant information, suggestively through replay. iEEG data of patients solving the same task shows that sharp-wave ripple rate is increased during this planning time. Together, our findings suggest an algorithm of how human mPFC encodes future actions, and provide evidence for a replay-based mechanism of loading that representation.

People are able to reason flexibly across a vast range of domains and contexts, from navigating new environments and social situations, to playing new games and even betting on the outcomes of new sports. How do we draw on our knowledge and past experiences to tractably make sense of any particular situation? Here, we explore the hypothesis that people use a combination of distributed and structured, symbolic knowledge to construct bespoke mental models tailored to novel situations. We propose a computational implementation of this idea -- a `Model Synthesis Architecture' (`MSA') -- using language models as a stand-in for distributional knowledge and a probabilistic programming language to express bespoke probabilistic symbolic models. We evaluate our model with respect to human judgments on a novel reasoning dataset. Our "Model Olympics" domain comprises a series of "sports commentary" vignettes, and is designed to test open-ended reasoning by requiring (i) reasoning about arbitrary causal structures described in language; (ii) drawing in relevant latent considerations from background knowledge; and (iii) flexibly adapting to an ‘open world' setting with novel observations sourced from other human participants. We compare our MSA to hand-coded probabilistic programs and LM-only baselines. We find that our approach captures key hallmarks of rational inference from human judgments that the LM-only baselines do not, especially for very novel scenarios. See https://sites.google.com/view/openworld-msa?usp=sharing for additional details and preprint.

Modal cognition is foundational to human reasoning, enabling us to construct and narrow a rich space of hypotheses about the world. While this faculty has often been considered uniquely human, its evolutionary roots remained uncertain. Nonhuman primates, lacking a natural language modal lexicon, provide a crucial test case for whether modal reasoning depends on linguistic expression. Prior research has cast doubt on primates' ability to contrast mutually exclusive possibilities, with subjects failing tasks that require reasoning about uncertain outcomes. Across four experiments, we show that rhesus macaques can reliably distinguish between certain and possible rewards, exhibiting unprecedented success on a 3-cup task. Primates' previous failures may stem not from an inability to reason modally but from the burden of representing absence as a possible alternative. By alleviating this burden, we uncover an early evolutionary footprint of modal reasoning that extends beyond humans and reveals an ancient, language-independent logical framework.

Social norms—perceptions of what is commonly done in a given context—serve as powerful signals for guiding moral decision-making in complex dilemmas. We investigate whether individuals adjust their moral strategies in response to information about others' judgments, and explore the underlying learning processes that support these shifts. Using a computational approach, we compare two models of social learning: (1) Decision Biasing, where social influence temporarily alters choices without affecting underlying values, and (2) Value Shaping, where social feedback directly updates individuals' moral value representations. Our results show that a Decision Biasing model fails to adequately explain the observed data, while a Value Shaping model better accounts for the persistence of moral adaptation. Taken together, these data suggest that social norms may play a role in shaping not only immediate moral choices but also the strategies people use to make them.

People conceptualize space using different spatial reference frames, based either in the body or the environment. Many studies attribute this cognitive diversity to spatial language, but their effects are confounded by differences across cultures and experimental tasks. Here we tested this hypothesis in people and tasks that are directly comparable. Indigenous BaYaka adults in the Congo basin reconstructed simple object arrays from memory and later described the same arrays aloud. Reference frames diverged across modalities in three primary ways. First, they used body-based frames seven times more often in their spatial memory than in their spatial descriptions of the same stimuli. Second, linguistic and non-linguistic responses were uncorrelated across participants, despite substantial variation. Third, each of two factors – spatial aspect and axis – influenced spatial language and memory differently, even in the same individuals. The results show that this fundamental feature of spatial thinking does not reflect patterns of spatial language.

The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models social learning as joint probabilistic inference of structured causal world models given both sensorimotor and linguistic data. Using behavioral and simulation studies of learning across 10 video games, we show how linguistic guidance shapes exploration by reducing risky interactions and speeding up key discoveries, in both humans and models. Most notably, we demonstrate successful cross-embodiment knowledge transfer: both human- and model-generated advice speeds up both human and model learning, revealing how structured, language-compatible representations might enable human-machine collaborative learning.

How cultural experience affects visual perception is a question of outstanding interest to debates regarding universality and cultural-specificity in human cognition. Yet, work comparing visual perception in 'typical' urban, industrialized samples with groups living in rural environments, typical for 99% of our species' history is strikingly limited (Deregowski, 2017). Here we more than double the total number of paradigms (visual illusions) in this literature, reporting data from 1) a 'typical' UK/US urban sample 2) a developing Namibian town 3) rural Namibian villages. Results reveal profound differences in visual processing, including aspects previously assumed to be universal (e.g. amodal completion in Gestalt shapes, line perception in the Cafe wall illusion). In contrast to recent arguments for the limited role of cultural experience in visual perception (Amir & Firestone, 2025), the present work indicates that a major research program in CCVS is warranted to capture the ways culture shapes visual perception.

Infants can understand some language even when they have no productive ability, but later in development, children can produce much more of what they understand. Explanations of this production--comprehension asymmetry (PCA) typically appeal to specific mechanisms, such as motor demands or communicative intent. Here, we investigate the hypothesis that the development of PCA emerges from the inherent structure of the two tasks. Production involves selecting a particular word to produce (and no other); in contrast, comprehension typically involves selecting the correct response to a word within a relatively constrained context. We tested this hypothesis by exploring whether developmental changes in PCA emerge in language models, which are sensitive to these structural asymmetries but not other factors previously proposed to cause PCA. We find that two types of language models---unimodal language models and vision–language models---both show PCA. Moreover, similar to children, PCA decreases in more highly-trained models and was more pronounced for predicates than nouns. These results suggest that production--comprehension asymmetries, a fundamental feature of child language acquisition, may be explained by the basic task demands involved in language use.

Understanding how cognitive and social mechanisms shape the evolution of complex artifacts such as songs is central to cultural evolution research. Social network topology (what artifacts are available?), selection (which are chosen?), and reproduction (how are they copied?) have all been proposed as key influencing factors. However, prior research has rarely studied them together due to methodological challenges. We address this gap through a controlled naturalistic paradigm whereby participants (N=2,404) are placed in networks and are asked to iteratively choose and sing back melodies from their neighbors. We show that this setting yields melodies that are more complex and pleasant than those found in the more-studied linear transmission setting, and exhibits robust differences across topologies. Crucially, these differences are diminished when selection or reproduction bias are eliminated in control studies, suggesting an interaction between mechanisms. These findings shed light on the interplay of mechanisms underlying the evolution of cultural artifacts.

Recent evidence suggests that Large Language Models (LLMs) display Theory of Mind (ToM) abilities. However, experiments with LLMs typically assess only *spectatorial* ToM, where LLMs merely predict other agents' behavior, rather than *planning*. In contrast, ToM in humans also contributes to dynamically *planning action* and *intervening* on others' mental states. We present a novel task of such a `planning theory of mind' (PToM), which requires agents to infer an interlocutor's beliefs and desires and persuade them to alter their behavior. We find that humans significantly outperform o1 (an LLM) at our task, even though o1 outperforms humans in a baseline condition which requires minimal mental state inferences. The results suggest that LLM performance at other ToM tasks may be attributable to simpler predictive abilities, while people excel at counterfactual planning when reasoning about others' behavior. Our paper is here: https://jaredmoore.org/mindgames

The ability to filter out distractors is crucial for effective sensory processing. Since the world is rarely random, prior knowledge can guide selective attention. Using a binocular rivalry paradigm combined with steady-state visually evoked responses, we investigated the neural dynamics of target and distractor processing. Behavioral enhancement was observed in the target-cueing condition, with no significant cost in the distractor-cueing condition. Single-trial analysis revealed that the distractor-related cost was mitigated by the complementary roles of parietal alpha and frontal theta. Specifically, prominent frontal theta activity during the rivalry phase indicated reactive control, reducing the distractor's sensory strength. Additionally, prior knowledge of distractors induced strong parietal alpha activity, reflecting pre-tuning of the attentional gate, which stabilized the target signal and blocked the distractor without affecting sensory gain. Our results reveal a dual mechanism for resolving visual competition, highlighting the distinct yet cooperative roles of parietal alpha and frontal theta.

Traditional research on children's emotion word comprehension has relied on explicit-response tasks and focused primarily on facial expressions, potentially underestimating early abilities. Using a looking-while-listening paradigm, this study examined whether 18- to 30-month-old children (N=100) could map emotion words to combined facial and bodily expressions. On each trial, children heard an emotion word while viewing a pair of emotional expressions that were either across valence (e.g., happy vs. sad) or within valence (e.g., sad vs. angry). Children aged 24-30 months preferentially looked at the matched expression on both trial types, while children aged 18-24 months old performed at chance levels. These findings suggest that the ability to map emotion words to facial and bodily emotional expressions emerges in early age two.

Homo sapiens maintain complex cooperative interactions with unrelated individuals by exploiting various cognitive mechanisms, for instance empathic reactions, the ability to tell cooperative agents from non-cooperative ones, a preference for prosocial individuals, and a desire for the punishment of antisocial individuals. The key role played by these features across moral systems suggests that core processes underpinning people's moral sense are evolved adaptations. Initial evidence consistent with this nativist view came from studies showing preferences for prosocial individuals in preverbal infants. Here we show that 5-day-old neonates can distinguish prosocial from antisocial interactions, looking longer at affiliative/helping behaviors than at avoidant/hindering behaviors. These visual preferences are specific to socially interactive stimuli, helping to rule out low-level perceptual explanations for the results. By revealing a preference for prosocial actions in newborns, these findings provide significant support for theories that posit evolutionary bases for at least some components of the human moral sense.

Languages are more likely to have lexical items for some concepts (e.g. CHILD) than others (e.g. PARENT). We propose that the communicative need of a concept influences how often it is lexicalized across languages, and test the hypothesis that usage frequency (which also reflects communicative need) predicts lexicalization across languages. Our analyses consider more than a thousand concepts, and demonstrate that average usage frequency across dozens of languages is a relatively good predictor of the typological prevalence of lexicalization across hundreds of languages. This finding implies that cross-linguistic regularities in lexicalization can be attributed in part to shared communicative need across cultures.

Humans can generalise from past experiences to novel situations as well as revise prior knowledge to flexibly adapt to changing contexts and goals. The representational geometry framework formalises how information is structured in the brain and suggests that abstraction involves a trade-off between generalisation and flexibility. However, how task representations evolve across learning and relate to behaviour remains unclear. Here, we tested the hypothesis that representational compression of task representations across learning underlies this flexibility impairment. Using an extra-dimensional shifting task, we manipulated the pretraining length to control the degree of compression. In both humans and artificial neural networks, longer pretraining was associated with decreased flexibility. Network dynamics indicated that greater compression incurred a higher representational reorganisation cost, limiting flexibility. Introducing an auxiliary reconstruction loss maintained higher dimensionality and mitigated the flexibility impairment. These findings suggest representational compression constrains flexibility, and preserving representational richness enhances flexibility.

One way to solve a repeated coordination problem is to generalize from past solutions: acting based on precedent or relying on existing rules. An alternative way is to reason about what would be optimally mutually beneficial in the moment. We investigate the trade-off between backward-looking behavior based on precedent and forward-looking reasoning about mutual benefit in a novel real-time incentivized coordination game (n = 252; 22,680 choices; preregistered). Then, we develop a cognitive model based on virtual bargaining and Bayesian inverse joint-planning which combines two components: one based on precedent, and one based on mutual benefit. Our model captures participants' behavior in the task, performs better than alternatives, and reproduces key differences between conditions in simulations.

Homing pigeons develop more efficient routes over repeated flights from a given location, but many open questions remain regarding the mechanisms behind this ability. This study examines the role of flight loops in navigation—instances where pigeons circle at specific locations during their homeward journeys. While loops are a source of navigational inefficiency, their navigational utility, if any, has not been studied. We adopt a data-driven approach to test hypotheses about looping on a GPS dataset of pigeons homing from a novel release site. We found that looping decreases with experience, with birds performing fewer and shorter loops with repeated releases. Less efficient navigators exhibited significantly more looping behavior. Our analysis revealed that directional uncertainty tends to decrease after loops compared to before, suggesting that loops may be an information-gathering mechanism. Additionally, locations where pigeons performed loops were more likely to be revisited in subsequent flights, indicating these sites might correspond to important and/or salient landmarks. Together, these findings illuminate the extent to which looping is not a purely stochastic navigational event, but a deliberate strategy.

Humans have the capacity for flexible and abstract reasoning that combines knowledge across distinct domains. To investigate the developmental origin of this capacity, we asked whether preverbal infants with limited language and no formal education can use logic to integrate information across physical and social domains. Across three preregistered experiments, we show that both adults and 14-month-old infants spontaneously use disjunctive reasoning to integrate predictions about possible outcomes of a probabilistic physical event with expectations about rational social behavior. In Experiment 1, adults deduced the agent's preference by eliminating alternatives. In Experiment 2, infants were surprised when the agent's behavior violated a logically inferred preference. Experiment 3 ruled out alternative novelty-based explanations. These findings suggest that infant logical inference can be reduced neither to domain-specific representations of physical objects or social agents, nor to language, revealing general-purpose core logic as a foundation of human thought.

Theory of Mind enables us to attribute mental states like beliefs and desires. We use it cooperatively, but we also use it adversarially, as when we lie. Prior work has shown people use Theory of Mind to craft lies to be believable to their audience, based on their audience's beliefs. But we usually also know something about our audience's desires. In this work, we ask a new question: Do people cater to their audience's desires by telling them what they want to hear? We propose that people expect others to be wishful thinkers---allowing their desires to color their beliefs---and exploit this by tailoring lies to audience desires. We implement this theory as a computational model and test it against human behavior in a novel task. This model quantitatively captures people's patterns of lying---both at the population and subject levels. This work advances our understanding of social cognition in adversarial interactions.

Understanding the relationship between seeing and knowing is fundamental to social cognition. While research demonstrates that even infants grasp basic aspects of this relationship, prior work often treats perceptual access and knowledge as equivalent (e.g., "if you see it, you know it"). In reality, their connection is richer: more complex objects require longer to encode, and agents' looking patterns often reveal how well they have encoded something and how much they want it. Across three experiments, we investigated whether children understand these nuances. In Experiment 1, we found that by age six, children expect more objects to require longer looking times. In Experiment 2, children inferred that agents who looked longer were more likely to form accurate representations of what they observed. In Experiment 3, children reasoned that agents who looked longer at an object were more likely to want it. Together, these findings suggest that by age six, children develop an intuitive theory of attention, enabling them to make sophisticated inferences about others' mental states based on looking behaviors.

Co-speech gestures arise from an interaction between visuospatial experience and speech formulation. Congenitally blind speakers produce gestures, but less than sighted speakers when describing spatial events. This study explores whether visual experience influences gesture kinematics to better understand the cognitive processes underlying gesture production. We conducted an auditory task where all participants listened to sounds of motion events (e.g., someone walking from a door). We analyzed co-speech path gestures (depicting the trajectory of the motion) spontaneously produced by 20 blind, 21 blindfolded, and 21 sighted Turkish speakers. We compared the alignment of speakers' path gestures with the actual spatial trajectory of the motions, along with other kinematic features—duration, size, and speed. Blind speakers took longer to produce larger gestures than sighted speakers. Blind speakers' gestures also reflected better precision than those of non-blind speakers—aligning with spatial cognition research. Thus, altered spatial cognition shapes gestures during event description.

A key feature of human theory-of-mind is the ability to attribute beliefs to other agents as mentalistic explanations for their behavior. But given the wide variety of beliefs that agents may hold about the world and the rich language we can use to express them, which specific beliefs are people inclined to attribute to others? In this paper, we investigate the hypothesis that people prefer to attribute beliefs that are good explanations for the behavior they observe. We develop a computational model that quantifies the explanatory strength of a (natural language) statement about an agent's beliefs via three factors: accuracy, informativity, and causal relevance to actions, each of which can be computed from a probabilistic generative model of belief-driven behavior. Using this model, we study the role of each factor in how people selectively attribute beliefs to other agents. We investigate this via an experiment where participants watch an agent collect keys hidden in boxes in order to reach a goal, then rank a set of statements describing the agent's beliefs about the boxes' contents. We find that accuracy and informativity perform reasonably well at predicting these rankings when combined, but that causal relevance is the single factor that best explains participants' responses.

It is well accepted in developmental and cognitive science that children represent the structure of their environment. This skill is generally useful, one that allows children to acquire the language(s) to which they are exposed, learn social rules, or represent relevant categories. But could this skill also have pernicious effects? Indeed, contemporary theories argue that children's racial biases emerge from being exposed to racial inequalities in their environment. Therefore, understanding – and ultimately interrupting – the development of racial biases requires better understanding the sources in young children's environment that perpetuate racial inequalities so that they can be corrected. Here we focus on a common aspect of children's environment – children's books – and document two features of how racial groups are depicted in U.S. children's books: the racial diversity (frequency) and racial representation (themes) in U.S. children's picture books. Through a meta-analysis (Study 1) and an information-theory analysis of an existing book collection (Study 2), we show that characters from minoritized racial backgrounds are not only numerically underrepresented in U.S. children's books, but there are also meaningful differences in the themes that are more likely to be associated with different racial groups. We discuss the implications of these results for developing theories of bias development that are grounded on real-world structure and for designing effective bias reduction approaches in childhood.

Contemporary transformer models have achieved human-like performance on many text-based tasks. However, real-world communication requires the integration of language with non-linguistic context (e.g., visual, social, etc.). Here, we study such information integration in three multimodal transformer models. We test these models' pragmatic capabilities regarding referring expressions: when an object set contains two exemplars from the same category that differ in size, unambiguously referring to one of them requires a size adjective (e.g., the big hammer); the adjective is unnecessary if only one exemplar from the category is present. We evaluate these inferences when models process text-image inputs (via their surprisal for infelicitous vs. felicitous adjective use) and when they generate open-ended descriptions of images given text prompts. We find evidence for pragmatic integration of visual and linguistic context in all models. However, these inferences remain sensitive to the in-context statistics of visual inputs, unlike pragmatic inference in humans.

Real-time human-artificial intelligence (AI) collaboration is crucial yet challenging, especially when AI agents must adapt to diverse and unseen human behaviors in dynamic scenarios. Existing large language model (LLM) agents often fail to accurately model the complex human mental characteristics such as domain intentions, especially in the absence of direct communication. To address this limitation, we propose a novel dual process multi-scale theory of mind (DPMT) framework, drawing inspiration from cognitive science's dual process theory. Our DPMT framework incorporates a multi-scale theory of mind (ToM) module to facilitate robust human partner modeling through mental characteristic reasoning. Experimental results demonstrate that DPMT significantly enhances human-AI collaboration, and ablation studies further validate the contributions of our multi-scale ToM in the slow system.

Humans can remember arrays or sequences of stimuli that exceed working memory limits in many domains, from auditory sequences to geometric shapes. This ability has been interpreted as evidence for language-like representations that compress stimuli into compact descriptions. We extend this evidence in the domain of geometry by showing that representations of geometric shapes are not only compressed but also syntactically structured. Experiment 1 shows that different representations can be induced for the same geometric shape, indicating structural representation. Experiment 2 shows that subparts of a shape are easier to recognize when they belong to the same subtree than whey do not, indicating hierarchical organization. Taken together, the results indicate that geometric shapes are encoded in representations that possess internal syntax, just like natural language sentences.

Conspiracy theories explain the cause of world events through the machinations of shadowy, secret groups. Understanding what in conspiracy theories makes them appealing explanations is an important area of research. People like explanations that can account for a large number of seen events (broad explanatory scope), while not accounting for every possible unseen event (narrow latent scope). It is unknown if people think conspiracy theories have broad or narrow scope and how this may relate to their believability. We thus explored perceptions of conspiracy theory explanations' scope and how this relates to their believability. Participants rated 40 conspiracy theories and their fact-based alternative explanations. Fact-based explanations were seen as having larger explanatory and latent scope. Additionally, larger scope was positively correlated with higher believability for both explanation types. We discuss how these findings relate to the explanation literature and highlight important elements of the seductive appeal of conspiracy theories.

We report a large study examining the effects of tDCS on the FIE. Subjects randomly assigned to one of the four tDCS groups and engaged with an old/new recognition task involving upright and inverted faces. 1) Sham tDCS during the study phase and recognition task; 2) Anodal tDCS during the study phase followed by sham tDCS during the recognition task; 3) Anodal tDCS during the study phase followed by cathodal tDCS during the recognition task; 4) Cathodal tDCS during the study phase followed by sham tDCS during the recognition task. Group 2 confirmed that anodal tDCS reduces the FIE vs. sham (Group 1) by disrupting performance for upright faces. Group 3 showed that cathodal tDCS applied after anodal, increased the FIE vs. Group 2, bringing it back to sham level, by enhancing upright faces. Group 4 revealed that cathodal tDCS applied after sham has no effect on the baseline FIE

Also known as the Ganong effect, a lexicality bias effect- i.e., the bias to interpret an ambiguous sound as the phoneme that yields a real word in its context - has been widely replicated. The search for a similar frequency bias effect, on the other hand, has yielded mixed results: In English, a bias has been observed such that listeners tend to interpret an ambiguous sound as the phoneme that yields a higher-frequency word, but this has failed to be replicated in Mandarin. One difference between these studies is the use of monosyllabic vs. disyllabic stimuli. To determine the factors that influence the presence of a bias effect, the present study tested for both frequency and lexicality bias effects using monosyllabic and disyllabic stimuli in Cantonese. Results show that the lexicality and frequency bias effects can be elicited in both monosyllabic and disyllabic stimuli, but the frequency effect is weaker.

Two experiments investigated how task demands influence sentence processing mechanisms, as reflected in neural responses at the disambiguating word. Participants read garden-path sentences with late-closure ambiguity (e.g., While the man hunted the deer ran into the woods) and answered comprehension questions while their brainwaves were recorded. Experiment 1 used standard questions (e.g., Did the man hunt the deer?), while Experiment 2 used "explicit" questions (e.g., Did the sentence explicitly say that the man hunted the deer?) to reduce inference-based responses. Results showed a typical P600 effect for ambiguous sentences in Experiment 1, indicating syntactic reanalysis. In Experiment 2, responses showed an N400 followed by Sustained Frontal Negativity, suggesting a shift toward plausibility evaluation and increased processing load. The explicit questions appear to alter underlying sentence processing mechanisms, prompting more effortful resolution of interpretive conflict beyond initial syntactic reanalysis.

Risk attitudes determine how decision-makers resolve tradeoffs when decisions involve uncertainty. The framing of a decision problem can affect these attitudes. Regulatory focus theory holds that (promotion-focused) frames that emphasize acquiring gains induce higher risk-taking than (prevention-focused) frames that emphasize avoiding losses. Here, we examine how this framing effect is moderated by individual differences in the internality of encoding style—the readiness to construe stimuli in terms of expectancies and pre-existing categories. In two experiments, participants could obtain costly information to aid their focal decision; thus, a risky choice corresponded to obtaining little or no information. Payoffs were framed to emphasize either gaining a bonus or retaining an endowed budget. The results of both experiments suggested that individuals with a more internal encoding style were more likely to be affected by the payoff framing. These results suggest that framings' effectiveness on risk-taking depends on individual differences in cognitive processing style.

Humans can categorize with only a few samples despite the numerous features. To mimic this ability, we propose a novel mixture of probabilistic principal component analyzers (mPPCA) model with dimension-reduced category representations, along with a theoretical analysis of rational dimensionality choices in categorization. Tests on the {\tt CIFAR-10H} natural image categorization dataset show that introducing a single principal component for each category effectively improves predictions of human categorization patterns. We further use mPPCA to account for human category generalization with very few samples. In our experiments with visual patterns of varying size and color, combining principal components and the hierarchical prior leads to significantly better predictions of human generalization within and beyond previously learned categories.

Semantic vectors derived from training on large text corpora (e.g., word2vec, BERT) are widely used as a methodological tool to model similarity of concepts. Recent work has demonstrated that a small amount of human training data can be used to fine-tune these vectors for modeling specific tasks. For example, human ratings of pairwise similarity can be used to estimate a set of dimensional weights, and these weights can improve estimates of human similarity ratings for held-out pairs. We applied this methodology to the semantic fluency task (listing items from a category) and find that category- specific weights can be used to identify the semantic category of a fluency list. The results have methodological implications for modeling retrieval in semantic fluency tasks, estimating semantic representations, and identifying semantic clusters and switches in fluency data.

Children are adept at learning the principles and properties governing their environment. However, this environment is often highly inconsistent: causes do not always bring about their effects; people do not always act according to their preferences. Past research shows that young causal learners readily reason from probabilistic evidence, but little is known as to how they reason about that evidence. This study presented preschoolers (N=114) with the behavior of three different causes—one consistently effective, one consistently ineffective, and one inconsistent—from one of three domains (social, mechanical, biological) and asked children to predict the future behavior of each. Children's predictions not only captured the different degrees of inconsistency observed in the evidence but also reflected differences in prior knowledge and expectations about inconsistency between domains. These results offer a novel, more nuanced look into early causal cognition and often-overlooked complexities of causal learning and reasoning in the real world.

Evidence from studies of knowledge revision show the refutation text is an effective tool for helping readers correct their inaccurate understandings. While refutation shows immediate benefits in group comparisons, it is unclear whether these benefits are restricted to particular retention intervals, if these benefits generalize across topics, or what predicts knowledge revision following refutation. In the present study, participants evaluated misconceptions across various science topics, then read a mix of refutation and expository texts and were prompted to rate their surprise and confusion after each text. Participants then re-evaluated the misconceptions immediately after reading the texts and again at a two-week delay. While both texts reduced misconception endorsement, refutation texts lead to a greater reduction in misconception endorsement on both the immediate posttest and at a two-week delay. Finally, exploratory analyses suggest participants' ratings of surprise predicted knowledge revision and partially mediated the effect of text type on knowledge revision.

Mobile phones provide an opportunity for a symbiotic interaction between humans and machines, which allows phones to collect human-centric data at anytime and anywhere. However, low-quality answers, which refer to the wrong answers, may be provided by users when they are asked excessive questions or in unsuitable contexts (e.g., driving). To solve this problem, we aim to design a methodology to collect more correct answers. We propose to use answer reaction time to annotate answer quality, to find a suitable number of daily questions, and the context factors that need to be considered according to their history records. We validated our methodology via the public dataset, which was collected by an extensive four-week in-the-wild study at the University of Trento, Italy. The results reveal that the context information and the number of daily questions are factors that can impact user answer behavior. These factors, therefore, influence the answer quality.

Alzheimer's disease (AD) is a highly heritable neurodegenerative disease, and brain imaging genetics (BIG) has become a key area for understanding its pathogenesis. However, existing methods often ignore the complex interrelationships between the multiple factors that lead to AD, especially when exploring the intrinsic connection between brain imaging features and gene variation. To address this challenge, we proposed a multimodal association modeling framework (MAM-GAN) based on generative adversarial networks, which aims to deeply reveal the association between genes and brain imaging features and apply it to disease state prediction. To verify the effectiveness of the framework, we conducted experiments using public datasets, and the results showed that MAM-GAN performed well in two classification tasks and successfully identified biomarkers closely related to AD.

Context-dependent reinforcement learning (RL) challenges the assumption that decision makers encode the absolute values of choice outcomes. This study investigates whether the associated choice biases arise from a relative encoding of outcomes or an alternative mechanism involving cumulative reward learning and selective attention to outcomes. Using eye tracking, participants completed a RL task where choice options were initially learned in fixed contexts before being tested in novel pairings. Results revealed an overall preference for options that were contextually favored in the learning phase, even when these preferences violated expected value maximization. Computational model comparisons demonstrated that hybrid encoding models, incorporating absolute and relative values, provided the best overall account of individual behavior. While eye fixations on choice outcomes decreased over trials, fixation-dependent RL models did not fit the data well, suggesting that overt visual attention patterns do not fully explain context-dependent choice biases.

When conducting experimental research, the research questions are often inherently linked (and limited) to the paradigm that is used. In this paper, we present a new experimental tool -- GRIS (Generating Representations in Space) -- that builds experiments where participants can manipulate objects on a screen. Through a series of three experiments on sentence acceptability, category typicality, and multi-dimensional similarity, we demonstrate how GRIS-based experiments allow cognitive scientists to approximate representational spaces for a variety of cognitive phenomena, expanding the set of possible research questions that cognitive scientists may ask.

Recent work in Bayesian, agent-based modelling of scientific communities has employed the Bala-Goyal framework to study the mechanisms involved when industry influence applies the so-called 'Tobacco Strategy' to undermine collective inquiry. Motivated by limitations of these models, we propose an alternative based on a recently introduced framework for normative argument exchange across networks. We implement representations of two distinct types of industry influence: `Obfuscating' influence directs inquiry to experiments with low expected value of information. `Misleading' influence filters private research and only communicates misleading signals from the world. We explored the impacts of both strategies on the polarization

The ongoing climate crisis demands massive changes in people's life style. Behavioral economics has highlighted the use of extrinsic incentives (e.g., money) as a powerful tool for changing behavior. However, external incentives come with significant costs, making them feasible primarily for wealthier countries. Here, following recent insights from the field of curiosity and information-seeking, we explore whether internal incentives such as the intrinsic drive for knowing can motivate people to act more pro-environmentally. By developing a novel decision-making task, we showed that the drive for knowing predicts pro-environmental choices. Moreover, participants chose eco-friendly options more when their values were unknown compared to when they were known. Results from this study hold the potential to inform the development of future behavioral interventions, although a replication of its findings is still ongoing.

"Let's share the load" is a much less helpful way to coordinate cleaning the house than "You take the bedrooms, I'll do the kitchen" - unless there are 12 bedrooms. Vague plans can be useful tools to support joint action, but using vagueness effectively is a difficult computational problem. Participants in a joint planning study selected between specific and vague plans to coordinate action across a systematic range of problems. In Experiment 1, participants deployed vague plans selectively, recognizing situations where the certainty of a bad plan is outweighed by the flexibility of a vague plan according to a probabilistic model of joint reasoning. In Experiment 2, participants with greater exposure to such situations endorsed vague plans when providing generic testimony to future actors. Our results highlight an understudied but potentially powerful dimension of human joint planning: strategic use of vague construals.

Classic research has found that young children are often egocentric when reasoning about others' visual experiences. In two experiments (total N = 148), we investigated 3- to 4-year-old children's abilities to reason about others' distinct visual experiences when they are engaged in social actions. Across experiments, we found that young children spontaneously oriented pictures and books so that those objects appeared upright to their social partners. These findings suggest that past research has underestimated young children's understanding of others' minds.

For more than two decades, researchers have been trying to explain the source of the processing cost of scalar implicature (SI). Although the computation of some SIs is associated with longer processing time (known as the delay effect), other SIs are processed cost-free. In this study, we investigated how individual differences in the rate of SI derivation modulate the delay effect across different scales. We reanalyzed four datasets from two SI verification task studies, which examined various scales. In these experiments, participants judged SI-triggering sentences as either true (literal reading) or false (SI reading). We fit a computational model to quantify the by-subject probability of computing SIs. Across datasets, we found that subjects who prefer the literal reading of the SI-triggering sentence were faster to respond true than false. However, the reading preferences modulate the verification speed differently for different scales. This suggests that the source of the delay effect might vary between scales.

This paper presents a formal account of causal explanation, grounded in a theory of conversational pragmatics, and inspired by the interventionist idea that explanation is about asking and answering what-if-things-had-been-different questions. We illustrate the fruitfulness of the account, relative to previous accounts, by showing that widely recognized "explanatory virtues" emerge naturally, as do subtle empirical patterns concerning the impact of norms on causal judgments. An extended version of the paper with further details can be found here: https://arxiv.org/pdf/2505.03732

The prepositions "in" and "on" are used in a wide variety of concrete, spatial contexts as well as abstract, non-spatial contexts. Previous research suggests that the meanings of these prepositions when used in abstract, non-spatial contexts might be grounded in the meanings these same prepositions have when they are used in concrete, spatial contexts; however, few studies have attempted to empirically test for evidence of these connections directly. The current study attempted to conceptually replicate Feist and Breaux (2013): No evidence of priming was observed. It could be that our prime stimuli did not adequately activate the meanings of "in" and "on" or that our participants' linguistic habits obscured any impact of the prime stimuli. Further research is needed to truly understand the role that grounded connections might play in the lexical semantics of words that are used in both abstract and spatial contexts.

How people reason about the mechanics of the physical world is an important question for several different cognitive sciences. Education, cognitive psychology, and developmental psychology have each conducted large numbers of studies over the last several decades, largely in isolation from one another (especially in the last quarter century). The results have suggested that cognitive mechanics may be subserved by a number of mechanisms that are differentially involved in different tasks. Here, we report converging results from factor analysis of a large compendium of mechanics questions.

Scientific representations and their constituent concepts change over time to reflect improvements in our understanding of the world. Similar improvements in understanding lead to changes in DNN-procured representations and their features. In this paper, we investigate whether useful methodological practices in concept change and in feature change carry across the two types of representations. We argue that there is indeed considerable potential for methodological cross-pollination and offer some examples of how such benefit may be derived.

Three experiments investigated preschoolers' ability to infer that numbers can be causally efficacious. Preschoolers observed that one of two quantities of objects activated a machine (i.e., a container holding 2 blocks activated a machine while a container holding 3 did not). Children were asked to determine whether novel containers with either 2 or 3 objects would activate the machine, and then construct their own container of objects that would do so. Four-year-olds, but not 3-year-olds, were above chance at both tasks given a contrast between the numbers ‘2' and ‘3' (Study 1), but not as good when the contrast was between the numbers ‘4' and ‘6' (Study 2) The effect of age on understanding ‘2' was mediated when children's numerical knowledge was considered (Study 3). These results are interpreted in terms of children's causal reasoning and hypothesis-formation abilities, but also their developing knowledge of numbers.

Recent legal rulings have denied copyright protection to artworks derived from AI-generated sources, because AI is assumed to be incompatible with qualities that define human authorship. We empirically test lay intuitions related to these assumptions in two studies (N = 235, N = 119) by investigating how creator attribution of initial source material (AI- vs. human-generated), effort investment in generating source material, and modification level of a derivative work influence perceptions of transformativeness, essence change, and creativity in derivative artworks. Modification level exerted the strongest influence across all measures, with dramatic modifications rated significantly higher than slight or no modifications. Effort investment in generating source material only influenced creativity ratings, with less effort sometimes perceived as more creative. Most notably, creator attribution for source material had minimal impact. These results challenge current copyright doctrine by demonstrating that lay human observers prioritize degree of transformation over both effort and creator attribution for source material. Our findings suggest that legal frameworks should recognize that AI assistance in generating artworks does not preclude a genuine human contribution that merits copyright protection.

This study introduces the tortoise attention algorithm (TAA), a novel computational tool for accurately measuring children's concentration during learning activities. The algorithm integrates weighted behavior duration and temporal stability metrics to calculate a comprehensive concentration score. We conducted three studies to evaluate the effectiveness of the algorithm. Study 1 demonstrated that TAA's concentration scores significantly predicted performance on math tasks but found no significant relationship with executive function performance, as measured by the ANT test. Study 2 revealed that TAA outperformed humans in predicting children's math performance, underscoring the algorithm's ability to mitigate biases inherent in human assessments of fidgeting behaviors. Study 3 further showed that while human evaluations of concentration were consistent across classroom and café settings, they failed to align with actual learning outcomes. These findings highlight that TAA provides an objective and reliable tool for evaluating concentration, enabling educators to refine teaching strategies and improve learning outcomes.

Social anxiety disorder (SAD) is characterized by cognitive biases that impair emotional information processing. This study examines how acute social stress affects working memory (WM) updating in socially anxious individuals using an emotional 2-back task. A 2×2 factorial design (N = 137) categorized participants into socially anxious (SA) and non-anxious (NSA) groups, further divided into stressed and control conditions. Acute stress was induced via speech anticipation, and subjective stress ratings were recorded. Mixed- effects modelling revealed independent negative effects of both social anxiety and acute stress on WM updating, including reduced accuracy, increased false alarms, lower discriminability (d-prime), and increased inverse efficiency score. Notably, disgusted facial expressions enhanced task efficiency under stress. Findings suggest stress-related cognitive deficits in social anxiety are additive rather than interactive, highlighting potential targets for intervention. This study contributes to understanding emotion-cognition interactions and extends research to understudied cultural contexts.

Mild cognitive impairment (MCI) detection using eye-tracking (ET) signals in virtual reality (VR)-based cognitive tasks shows great promise, as it can capture rich temporal and behavioral information. Therefore, we build four VR-based tasks based on Stroop test and construct a dataset for MCI detection using ET signals. However, ET signals often suffer from non-stationarity,variability, and redundancy, challenging accurate MCI detection.To address these issues, we propose a novel lightweight network KWS-TA-CNN with three key components: 1) Kymatio Wavelet scattering transform (KWS), which generates time-robust features and reduces memory usage through a depth-first traversal strategy; 2) Temporal Attention (TA) to dynamically weight critical time steps for MCI detection; and 3) 1D Convolutional Neural Network (CNN) to capture local temporal patterns and reduce feature redundancy. Experimental results from leave-one-subject-out cross-validation show high performance, with subject-level accuracies of 0.8158, 0.9211, 0.8158, and 0.8421 across the four tasks, demonstrating its strong clinical potential.

Emotion recognition based on electroencephalography (EEG) plays a significant role in brain-computer interface (BCI) applications. However, individual differences often hinder the generalization of emotion recognition methods to unknown subjects. To address this, we propose an unsupervised domain adaptive model with joint multi-source domain and two-step alignment strategy (JMS2A). The alignment strategy consists of two steps: (1) To capture the structured information from source domain, we combine multiple source domains into a mixed source domain. Simultaneously, a single source domain and the target domain are combined to form a pseudo-target domain, which is then indirectly aligned with the mixed source domain; (2) To extract latent class information from the target domain, we extend supervised contrastive learning to enable direct alignment between source and target domain. We evaluated JMS2A on the SEED and SEED-IV datasets, achieving accuracies of 95.30% and 86.55%, respectively. Experimental results demonstrate that our approach outperforms state-of-the-art methods. The source code is available at https://github.com/cccyangt/JMS2A.

Language encodes societal beliefs about social groups through word patterns. While computational methods like word embeddings enable quantitative analysis of these patterns, studies have primarily examined gradual shifts in Western contexts. We present the first large-scale computational analysis of Chinese state-controlled media (1950-2019) to examine how revolutionary social transformations are reflected in official linguistic representations of social groups. Using diachronic word embeddings at multiple temporal resolutions, we find that Chinese representations differ significantly from Western counterparts, particularly regarding economic status, ethnicity, and gender. These representations show distinct evolutionary dynamics: while stereotypes of ethnicity, age, and body type remain remarkably stable across political upheavals, representations of gender and economic classes undergo dramatic shifts tracking historical transformations. This work advances our understanding of how officially sanctioned discourse encodes social structure through language while highlighting the importance of non-Western perspectives in computational social science.

The onset of spontaneous thoughts are reflective of dynamic interactions between cognition, emotion, and attention. Typically, these experiences are studied through subjective appraisals that focus on their triggers, phenomenology, and emotional salience. In this work, we use linguistic signatures to investigate Déjà Vu, Involuntary Autobiographical Memories, and Unexpected Thoughts. Specifically, we analyze the inherent characteristics of the linguistic patterns in participant generated descriptions of these thought types. We show how, by positioning language as a window into spontaneous cognition, existing theories on these attentional states can be updated and reaffirmed. Our findings align with prior research, reinforcing that Déjà Vu is a metacognitive experience characterized by abstract and spatial language, Involuntary Autobiographical Memories are rich in personal and emotionally significant detail, and Unexpected Thoughts are marked by unpredictability and cognitive disruption. This work is demonstrative of languages' potential to reveal deeper insights into how internal spontaneous cognitive states manifest through expression.

As web media, such as social networking services (SNS), become more prevalent, the formation of false beliefs through fake news and propaganda has become a significant problem. This study focuses on the cognitive process of users as actively information-seeking agents in web media exploration and proposes WEB-FEP, a computational model of users forming specific beliefs through interactions with web media. WEB-FEP specifically attempts to computationally reproduce confirmation bias in web media exploration by formalizing the trade-off between belief-confirmatory and exploratory actions inspired by active inference. WEB-FEP is validated by comparing the results of simulations with user experiments conducted on a virtual SNS. The results indicate that the initial belief distributions and learning rates modeled in WEB-FEP can successfully reproduce the diverse behaviors of users including confirmatory exploration.

The cognitive mechanism by which Large Language Models (LLMs) solve mathematical problems remains a widely debated and unresolved issue. Currently, there is little interpretable experimental evidence that connects LLMs' problem-solving with human cognitive psychology. To determine whether LLMs possess human-like mathematical reasoning, we modified the problems used in the human Cognitive Reflection Test (CRT). Our results show that even with the use of Chain-of-Thought (CoT) prompts, mainstream LLMs, including the o1 model (noted for its reasoning capabilities), have a high error rate when solving these modified CRT problems. Specifically, the average accuracy rate dropped by up to 50% compared to the original problems. Further analysis of LLMs' incorrect answers suggests that they primarily rely on pattern matching from their training data, which aligns more with human intuition (System 1 thinking) rather than with human-like reasoning (System 2 thinking). This finding challenges the belief that LLMs have genuine mathematical reasoning abilities comparable to humans. As a result, this work may adjust overly optimistic views on LLMs' progress toward Artificial General Intelligence.

Research on how humans perceive aesthetics in shapes, colours, and music has predominantly focused on Western populations, limiting our understanding of how cultural environments shape aesthetic preferences. We present a large-scale cross-cultural study examining aesthetic preferences across five distinct modalities extensively explored in the literature: shape, curvature, colour, musical harmony and melody. We gather 401,403 preference judgements from 4,835 participants across 10 countries, systematically sampling two-dimensional parameter spaces for each modality. The findings reveal both universal patterns and cultural variations. Preferences for shape and curvature cross-culturally demonstrate a consistent preference for symmetrical forms. While colour preferences are categorically consistent, ratio-like preferences vary across cultures. Musical harmony shows strong agreement in interval relationships despite differing regions of preference within the broad frequency spectrum, while melody shows the highest cross-cultural variation. These results suggest that aesthetic preferences emerge from an interplay between shared perceptual mechanisms and cultural learning.

As Large Language Models (LLMs) become more prevalent, concerns about their safety, ethics, and potential biases have risen. Systematically evaluating LLMs' risk decision-making tendencies, particularly in the ethical domain, has become crucial. This study innovatively applies the Domain-Specific Risk-Taking (DOSPERT) scale from cognitive science to LLMs and proposes a novel Ethical Decision-Making Risk Attitude Scale (EDRAS) to assess LLMs' ethical risk attitudes in depth. We further propose a novel approach integrating risk scales and role-playing to quantitatively evaluate systematic biases in LLMs. Through systematic evaluation of multiple mainstream LLMs, we assessed the "risk personalities" of LLMs across multiple domains, with a particular focus on the ethical domain, and revealed and quantified LLMs' systematic biases towards different groups. This helps understand LLMs' risk decision-making and ensure their safe and reliable application. Our approach provides a tool for identifying biases, contributing to fairer and more trustworthy AI systems.

We examined whether embodiment effects at a particular word influence subsequent words. We recorded EEG while participants read action sentences that were literal-concrete (LC), literal-abstract (LA), and metaphorical (MET). Prior work showed that at the verbs, both LC and MET elicited more negative N400s than LA, reflecting sensorimotor simulations. We found that at the object nouns, LC elicited a more negative negativity in the 500-700 ms time window than LA and MET, which may reflect combined influences from embodiment spilled over from the verbs and imagery linked to noun concreteness and imageability. These findings suggested that literal action embodiment can yield extended motor engagement extending into subsequent words, but metaphorical action verbs cannot. Moreover, lexical properties at the noun played a role.

Past methods for studying graph interpretation have only indirectly assessed people's mental picture of the data that produced the graph. Recently, we developed a more direct, drawing-based measure and used to reveal a severe misinterpretation of the common bar graph of averages: one in five viewers mistook the average for the data's outer limit. Here, we use the same measure to reveal a second misinterpretation, whereby even more viewers—one in three—incorrectly assume that data frequency remains approximately uniform over its entire range. Missing from their mental picture are the tails of the distribution—the relative rarity of extreme values—which are so characteristic of real data that they are embedded in the core normality assumption of statistics. We label this misinterpretation the "Uniformity Fallacy" and characterize its nature, reproducibility, generalizability, and correlates. We conclude that bar graphs of averages fail to communicate data truthfully in not one, but two fundamental ways.

In human-AI decision-making, understanding the factors that maximize overall accuracy remains a critical challenge. This study highlights the role of metacognitive sensitivity—the agent's ability to assign confidence scores that reliably distinguish between correct and incorrect predictions. We propose a theoretical framework to evaluate the impact of accuracy and metacognitive sensitivity in hybrid decision-making contexts. Our analytical results establish conditions under which an agent with lower accuracy but higher metacognitive sensitivity can enhance overall decision accuracy when paired with another agent. Empirical analyses on a real-world image classification dataset confirm that stronger metacognitive sensitivity—whether in AI or human agents—can improve joint decision outcomes. These findings advocate for a more comprehensive approach to evaluating AI and human collaborators, emphasizing the joint optimization of accuracy and metacognitive sensitivity for enhanced decision-making.

The increasing adoption of conversational agents powered by large language models (LLMs) raises questions about its effects across culturally diverse interactions. While these agents are linguistically versatile and multilingual, their ability to adapt along cultural dimensions–defined as geographically and communally nurtured sets of values and behavioral norms– lacks close scrutiny of both their design and deployment. To achieve inclusive conversational AI, it is essential to understand how agents adapt to users from diverse cultural backgrounds. In this study, we analyze dialogues between human users from different countries and LLM-powered agents to examine how both parties adapt their word use, a salient aspect of linguistic styles, toward one another throughout casual conversations. Our analysis reveals that LLMs exhibit varying degrees of style matching based on users' national cultures and demonstrate asymmetric adaptation when interacting with culturally diverse users. Moreover, we observe a reciprocal dynamic where both the LLMs and users from certain cultures adjust their styles in response to one another. Additionally, our findings support the hypothesis that LLMs and users naturally converge in conversational styles over the course of interactions, mirroring the dynamics of human conversations that accommodate and converge. To develop localized and culturally aware agents, there's a potential to utilize such cross-cultural convergence process during fine-tuning to align LLMs.

Past research on eye-movements during reading and comprehension has primarily focused on alphabetic scripts, such as the Roman script used to write European languages like English, Dutch, German, and Spanish, where classical measures like word length can be easily calculated by counting characters. However, this approach may not generalize to alphasyllabic languages like Hindi and other Indian languages written using the Devanagari script, where many characters depend on diacritic markers for proper pronunciation. This poses challenges in studying these languages in eye-tracking research, discourages eye-tracking studies with these languages. To address this gap, we asked 61 native Hindi speakers (L1) read Hindi text, while their eye-movements were being tracked. Results revealed that a complexity metric for the script predicts variables such as first fixation duration, gaze duration, single fixation duration, total reading time, and number of fixations. These results were also correlated with variables such as word frequency (van Heuven et al., 2014) for all eye-tracking measures.

Humans recognize social interactions effortlessly, even when presented with minimal visual information in unfamiliar displays. While force dynamics has been proposed as latent representations for perceiving social interactions, most research has approached this topic from a linguistic perspective based on conceptual knowledge, leaving open the central question of how latent force representations arise from visual inputs. The present study developed a force model that represents social interactions through two types of compositional forces: interactive forces, driven by interactions between agents; and self-propelled forces, driven by intentions of individual agents. Each force was formulated using a physics function to capture the dynamics of repulsive and attractive forces. We conducted two human experiments to measure human similarity judgments across a range of interaction animations and to evaluate recognition performance using generated animations in which the forces applied to individual agents were systematically manipulated. We found that the force model provides a parsimonious account for human judgments in both experiments. These findings suggest that mid-level representations based on compositional forces driven by different goals play an important role in social perception. We conjecture that the development of social perception may be grounded in perceptual mechanisms that support intuitive physics.

Recent research in urban visual analytics has focused on crowdsourcing judgements of qualities such as beauty and safety, for large-scale city-wide predictive evaluation. This study examines the extent and nature of inter-subject variability in such judgements and argues that these subjective qualities often have low generalizability across individuals. We conducted an online study involving 94 participants across 19 countries, where subjects arranged streetscape scenes on visual scales of beauty and safety. An analysis of the arrangements revealed very low inter-subject consistency, including within demographic groups based on age, sex, race and nationality. K-means clustering also revealed large clusters with contradictory judgements with respect to visual features. There was however higher intra-subject consistency when rating the same scenes twice. Based on these findings, we recommend a cautious approach to the use of "average" crowdsourced judgements of urban qualities, and encourage the adoption of subject-specific prediction when evaluating such qualities at scale.

Electrical status epilepticus during sleep (ESES) is a serious condition that causes notable cognitive decline. It is characterized by distinct spike and slow-wave patterns on electroencephalograms (EEG). Clinical ESES diagnosis is extremely time-consuming and labor-intensive as it demands clinicians to manually interpret and count EEG screens. Existing automated diagnosis algorithms for ESES have major flaws, like struggling to adapt to complex spike-and-wave patterns and not fully exploiting the rich multi-view features of EEG. To overcome these issues, we propose a multi-view feature selection framework integrating reinforcement learning and attention mechanisms for automated ESES diagnosis. A CLEAN reward mechanism is introduced to address complex multi-objective feature selection challenges. Experiments on the clinical data consisting of 36 epilepsy patients prove the proposed method's remarkable spike-and-wave identification ability and high agreement with expert diagnoses. Our approach represents a significant step toward developing automated bedside ESES clinical diagnostic systems.

In the present study, participants learned the locations of 12 landmarks by following a guided route either in a desktop virtual environment or on an omnidirectional treadmill paired with a virtual reality headset. Spatial learning and wayfinding efficiency were later assessed across both interface conditions. Additionally, the ratio of physical to cognitive costs for navigating to goals was manipulated across trials. Results indicated that although the two groups did not differ in spatial learning, participants navigating on treadmills selected more efficient routes than those in the desktop group in trials involving high physical cost. Higher levels of self-report cybersickness were associated with reduced spatial learning and wayfinding efficiency, independent of interface condition. These findings validate the use of omnidirectional treadmills for investigating the tradeoff between cognitive and physical effort in navigation. At the same time, reducing cybersickness is essential to ensure the effective use of this technology.

Existing theories of word learning largely focus on a learner's ability to learn a single meaning for a word despite the fact that many words have multiple meanings. Several computational models of cross-situational word learning have been proposed to explain how words are learned, but it is unknown to what extent they can learn ambiguous words with multiple meanings. Here, we present an experiment showing that adult learners are able to learn multiple meanings of novel ambiguous words in a cross-situational word learning paradigm, and are especially good at doing so when the meanings of the words are related (polysemous) rather than unrelated (homophonous). We evaluated the ability of ten different computational models of cross-situational word learning to explain the empirical data, and none were able to learn the ambiguous words as successfully as the adult learners. Moreover, because these computational models do not represent any semantic information, they are in principle unable to replicate the key difference between polysemous and homophonous word learning found in the study.

Prevalence-induced concept change (PICC) occurs when reduced category prevalence increases the likelihood that ambiguous stimuli are classified as belonging to the now-minority category. PICC has been observed across perceptual and social domains and persists despite instructions or incentives to suppress it. However, other findings suggest its expression is instead shaped by social context. If AI models are to be treated as theories of cognition, they should exhibit PICC as well. We show that a standard AI model for sequential learning, trained on dynamic category distributions, does not display PICC, instead favoring the now-majority category. This opposite-PICC effect suggests that simple sequential learning may be insufficient to produce PICC. Additional mechanisms, such as structured priors, contextual sensitivity, or internal feedback, seem necessary for its emergence. Our findings contribute to the understanding of PICC and its implications for categorization theories, AI-driven decision-making, and the role of media exposure in shaping social perceptions.

Causal reasoning is a core component of intelligence. Large language models (LLMs) have shown impressive capabilities in generating human-like text, raising questions about whether their responses reflect true understanding or statistical patterns. We compared causal reasoning in humans and four LLMs using tasks based on collider graphs, rating the likelihood of a query variable occurring given evidence from other variables. LLMs' causal inferences ranged from often nonsensical (GPT-3.5) to human-like to often more normatively aligned than those of humans (GPT-4o, Gemini-Pro, and Claude). Computational model fitting showed that one reason for GPT-4o, Gemini-Pro, and Claude's superior performance is they didn't exhibit the "associative bias'' that plagues human causal reasoning. Nevertheless, even these LLMs did not fully capture subtler reasoning patterns associated with collider graphs, such as "explaining away". These findings underscore the need to assess AI biases as they increasingly assist human decision-making.

This paper presents a methodology combining multimodal semantic analysis with an eye-tracking experimental protocol to investigate the cognitive effort involved in understanding the communication of future scenarios. We conduct a pilot study examining how visual fixation patterns vary during evaluation of valence and counterfactuality in fictional ad pieces describing futuristic scenarios, using a portable eye tracker. Participants' eye movements are recorded while evaluating the stimuli and describing them to a conversation partner. Gaze patterns are analyzed alongside semantic representations of the stimuli and participants' descriptions, constructed from a frame semantic annotation of both linguistic and visual modalities. Preliminary results show that far-future and pessimistic scenarios are associated with longer fixations and more erratic saccades, supporting the hypothesis that fractures in the base spaces underlying interpretation of future scenarios increase cognitive load for comprehenders.

Understanding how children's spontaneous language behavior relates to standardized metrics of language development remains a crucial challenge in developmental science, particularly given the time and resources required for many traditional lab-based assessments. This study investigates whether automated analysis of naturalistic, child-centered audio recordings can index the developmental trajectory of speech-language abilities. Using a longitudinal design following N=130 preschoolers, we employed deep learning methods to compute Canonical Proportion - a theoretically-motivated metric that reflects both speech motor control development and phonological representation building - from naturalistic, child-centered audio recordings at age 3 years. Canonical proportion measures significantly predicted multiple dimensions of speech-language development longitudinally, formally assessed in the lab one year later at age 4. The strongest relationships were found for consonant articulation skill and vocabulary size, suggesting that early speech production patterns may moderately index numerous later facets of language development. These findings outline a potential relationship between children's spontaneous, everyday language behavior and more traditional language development metrics, while demonstrating the potential for automated measures to expand and diversify research in developmental science.

Human memory exhibits vulnerability in cognitive tasks; comparing visual working memory with new perceptual input can cause unintended distortions. Prior studies report systematic memory distortions post-comparison, but understanding their impact on real-world objects and identifying contributing visual features remains challenging. We propose an AI-driven framework generating naturalistic stimuli based on behavioral object dimensions to elicit similarity-induced memory biases. Using two stimuli types—image wheels (dimension-edited) and dimension wheels (activation-based)—we conducted three visual working memory experiments under conditions: no perceptual comparison, image wheel comparison, and dimension wheel comparison. Results show that both similar images and dimensions induce memory distortions. Visual dimensions (e.g., shape/texture) are more distortion-prone than semantic ones (e.g., category), indicating that naturalistic stimuli's object dimensions critically influence memory vulnerability.

Spatial navigation is a promising cognitive marker for several cognitive disorders. However, its clinical utility is still limited because spatial abilities are also influenced by non-pathological factors. Among these, the Cognitive Reserve (CR) stands out due to its strong influence on cognitive aging trajectories. However, the association between CR and spatial navigation across the lifespan has never been assessed. In this study, we collected spatial navigation data along with CR measures in a population with diverse demographic profiles. We found a strong decline in spatial ability with age, but no overall effect of the CR. We systematically analyzed the association between each CR item and spatial ability and found that the subscore related to reading habits was significantly associated with spatial navigation performance, even after correcting for multiple comparisons. We discuss the implications of these findings for early and personalized screening for cognitive disorders using a spatial navigation task.

Self-regulated learning (SRL) significantly influences students' learning behaviors and academic performance. However, research has focused on "between-person" differences, neglecting "within-person" variability. Traditional SRL assessments rely on self-reports, which fail to capture fine-grained behavioral changes (such as hourly variations). We propose a novel approach using mobile sensing to assess SRL through within-person variability. We use passive sensing data from the phones of 211 university students to explore this relationship. To assess behavioral variability, we focus on five sensing behaviors—physical activity, social interactions, sleep, location, and app usage data—and calculate four within-person variability features: standard deviation, circadian rhythm, regularity index, and flexible regularity index. Our findings reveal significant associations between these variability features and self-reported SRL skills, particularly in dimensions such as environment structuring, time management, and help seeking. This research provides new insights into assessing SRL and offers a theoretical foundation for future personalized interventions in educational settings.

The mapping between nonsymbolic quantities and symbolic numbers lays the foundation for mathematical development in children. However, the neural mechanisms underlying this crucial cognitive bridge remain unclear. Here, we in- vestigate the computational principles governing symbolic- nonsymbolic integration using a biologically inspired neural network trained through developmentally inspired stages. Our investigation reveals that generalization from nonsymbolic to symbolic numerical processing emerges specifically when rep- resentational alignment forms between these numerical for- mats. Notably, this alignment appears to be stronger in cross- format comparison-based mapping compared to direct-label- based mapping. Furthermore, we demonstrate that subsequent symbolic specialization creates a representational divergence that impairs nonsymbolic performance while maintaining the ordinal structure of the mapping. These findings highlight rep- resentational alignment as a fundamental mechanism in nu- merical cognition and suggest that targeted cross-format com- parison tasks may be particularly effective in improving math- ematical learning in children with numerical processing diffi- culties. Keywords: Emergence of number semantics, Representa- tional alignment, Artificial neural network

While modern computational models accurately predict free viewing scanpaths via learned deep neural networks, the in- vestigation of the brain's implementation of the modeled be- havior has been thus far limited to primarily bottom-up vi- sual features. One of the key question in the investigation in the early visual system's implementation of bottom-up visual saliency is how (and whether) it dynamically adapts its sensi- tivity to features of different spatial scales. This paper provides a simple test of whether identical stimuli presented at different spatial scales produce different gaze behavior. We asked sub- jects (n=12) to freely view video stimuli twice each session over two sessions In one session (intervention) the visual scale changed from large (25 degrees of visual angle) to small (10 degrees) between viewings. In the other session (control) the video size was the same. Gaze was more strongly correlated between viewings of the same video size (r=0.265) than differ- ent sizes (r=0.231), independent of whether there was a long (> 24 hours) or short (< 10 min) delay between viewings, im- plying that memory effects are not a strong factor. Although low, these within-subject correlations are higher than the cor- relation of gaze between different subjects viewing identical videos (r=0.195).

AI models are often developed to solve reasoning problems optimally. In contrast, cognitive models focus on explaining and predicting replicative cognitive patterns of human information processing. And while many of the theories aim to explain an assumed ‘general' human reasoner, only few are aimed at the individual. This paper addresses the challenge of the latter by investigating the automatic generation of individualised predictive algorithms using transformer-based models. These models which have been trained on huge amounts of human data, potentially exhibit built-in cognitive patterns. Leveraging such characteristics and architecture of transformer-based models, we outline a generalized methodology for establishing a human-AI collaborative framework, to generate explainable and reproducible algorithms with cross-domain applicability. While predictive accuracy and generalizability pose less of a problem, the bigger challenges in using machine learning approaches or transformer-based models may be explainability and replicability. Hence, instead of ‘just' using such a model for directly fitting the data, we use it to extract features and to propose cognitive algorithms that are executable in systems outside of the model. Using two datasets pertaining to syllogistic and spatial reasoning, the predictive algorithms thus generated applying the presented framework, achieve mean accuracies of 68% and 81%, respectively. Both algorithms outperform other established, state-of-the-art cognitive models by far, surpassing the (previously) best state-of-the art models in syllogistic and spatial human reasoning by 19% and 13%, respectively.

Recent advances in computational modeling have expanded our capacity to analyze language and communication, particularly through transformer models. The present work investigates how such computational frameworks can be leveraged to address clinical domains in communication disorders. We used semantic embeddings from BERT's layers to analyze language-related adjustments used by participants with traumatic brain injury (TBI) in conversational transcripts. By examining semantic convergence patterns across different layers of the BERT model, we found that TBI participants demonstrated more pronounced "self" convergence -- they tended to stay closer to their own semantic contributions in the conversation -- compared to controls. This effect was particularly noticeable at earlier layers of the BERT model, suggesting that surface-level semantics play a significant role. The findings highlight the potential for language models to enhance our understanding of social interaction dynamics. We further discuss how bridging computational linguistics with clinical domains can address analytic challenges in the study of natural cognition and communication.

The Tritone Paradox is an auditory illusion in which a sequence of two complex tones is perceived as either ascending or descending, depending on the individual. It presents an interesting phenomenon for investigating pitch perception in contexts. however, no neurophysiological study has been conducted. This study identified event-related potential (ERP) correlates of pitch judgments under different pitch contexts. Twenty-seven participants judged whether the tritone pair was perceived as ascending or descending after listening to a sequence of ascending or descending tone pairs. Cortical auditory evoked responses to the second tone of the tritone pair were compared across contexts. In the Rise context, standard stimuli evoked larger responses at Fp1; in the Fall context, deviant stimuli elicited stronger responses across all sites. These results suggest that frontal and central brain regions are involved in processing ambiguous pitch stimuli, and that ERP responses reflect the interaction between stimulus context and perceptual.

Accurate recognition of human emotions from EEG signals plays a critical role in affective computing and human-computer interaction. However, existing methods face significant challenges in effectively capturing the sparse, dynamic, and energy-dependent characteristics of brain activity during emotional experiences. To address these challenges, we propose a novel framework, Spatial-Energy-Aware Dynamic Filtering with Sparse Graph Convolutions (SEASGC), which rethinks EEG graph modeling from three perspectives: (1) sparse graph construction to adaptively capture the essential functional relationships between brain regions, (2) dynamic and location-dependent filtering to model nonlinear interactions between EEG nodes, and (3) energy-aware feature aggregation to leverage energy changes as critical indicators of emotional intensity. By explicitly integrating these principles, SEASGC provides a more comprehensive representation of EEG signals for emotion recognition. Extensive experiments on benchmark EEG emotion datasets demonstrate that SEASGC achieves state-of-the-art performance, highlighting its effectiveness and generalizability in modeling the complex spatial-spectral dynamics of EEG signals.

Computational models of human learning can play a significant role in enhancing our knowledge about nuances in theoretical and qualitative learning theories and frameworks. There are many existing frameworks in educational settings that have shown to be verified using empirical studies, but at times we find these theories make conflicting claims or recommendations for instruction. In this study, we propose a new computational model of human learning, Procedural ABICAP, that reconciles the ICAP, Knowledge-Learning-Instruction (KLI), and cognitive load theory (CLT) frameworks for learning procedural knowledge. ICAP assumes that constructive learning generally yields better learning outcomes, while theories such as KLI and CLT claim that this is not always true. We suppose that one reason for this may be that ICAP is primarily used for conceptual learning and is underspecified as a framework for thinking about procedural learning. We show how our computational model, both by design and through simulations, can be used to reconcile different results in the literature. More generally, we position our computational model as an executable theory of learning that can be used to simulate various educational settings.

Brain-to-Image reconstruction aims to recover visual stimuli perceived by humans from brain activity. However, the reconstructed visual stimuli often missing details and semantic inconsistencies, which may be attributed to insufficient semantic information. To address this issue, we propose an approach named Fine-grained Brain-to-Image reconstruction (FgB2I), which employs fine-grained text as bridge to improve image reconstruction. FgB2I comprises three key stages: detail enhancement, decoding fine-grained text descriptions, and text-bridged brain-to-image reconstruction. In the detail-enhancement stage, we leverage large vision–language models to generate fine-grained captions for visual stimuli and experimentally validate its importance. We propose three reward metrics (object accuracy, text-image semantic similarity, and image-image semantic similarity) to guide the language model in decoding fine-grained text descriptions from fMRI signals. The fine-grained text descriptions can be integrated into existing reconstruction methods to achieve fine-grained Brain-to-Image reconstruction.

Spoken sentence processing often requires integrating the linguistic message with information about the talker and the social context. Among other things, information about the talker's gender identity could influence how a listener processes what they hear, due to the prevalence of gender-related stereotypes in human societies. Several previous studies showed that listeners anticipated stereotype-consistent content, and that comprehension was affected when gender stereotypes were violated (e.g., when women serve in conventionally male-dominant professions, or vice versa). However, the existing findings are rather mixed. In this study, we examine the influence of talker gender information on language comprehension and prediction in Mandarin Chinese. We report the results of a cloze task where participants were asked to guess the last word of a sentence with or without information about talker gender. When talker gender information was available, we further varied whether gender information was revealed by personal names or voices. Participant responses were evaluated for gender bias by two pre-trained language models based on Word2Vec and GPT2. Results of statistical analysis revealed that participants adjusted their responses to align with the gender category cued by the sentence (i.e., more male/female-biased responses when the sentences implicated a male/female talker), but the effect was only present when gender information was implicated through names but not voices. The current study provides partial evidence for the effect of talker gender on sentence prediction. We discuss the implications of current findings for further study of the integration of linguistic and social information in language comprehension.

In languages which mark both gender and number as distinct morphemes, there is a tendency to place gender closer to the noun stem than number. However, the typological data on this is sparse. Moreover, linguistic theories differ in how they explain ordering patterns of gender and number morphology: some theories focus on the structure of the representations of features in the speakers' minds, and other focus on the role of co-occurrence statistics. In a recent study, Saldana, Kanampiu, and Culbertson (2025) use artificial language learning to show that learners with a diverse range of language experience with grammatical gender and number exhibit a consistent bias for orders with gender closer to the noun stem than number. This order reflects the ordering in which most linguistic theories assume number and gender features are derived in word formation. Here, we build on this study to investigate how this bias interacts with the statistics of the linguistic input. In particular, we manipulate co-occurrence between stems and affixes so that learners are exposed to combinations of stems and number morphology more often than to stems and gender morphology. We test whether input statistics can push learners to reverse their natural preference, leading them to place number closer to the noun than gender. We find that our manipulation reduces, but does not eliminate or reverse, the preference for gender-closest order. However, our study also highlights some difficulties learners have in acquiring novel features from sparse data. Ultimately, our findings highlight the dynamic interplay between representations of meaning and input-based learning mechanisms.

Emotion recognition is a critical area in brain-computer interfaces, with electroencephalography (EEG) shown to be effective for emotional analysis. In domain generalization, cross subject emotion recognition encounters significant generalization challenges, including excessive feature collapse and insufficient capture of EEG features. To tackle these issues, we propose an Anti-Collapse Cross-Domain Consistency Network (AC-CDCN), which leverages Maximum Mean Discrepancy (MMD) to reduce distribution discrepancies between source domains, facilitating the capture of domain-invariant features, and innovatively introduces an Anti-Feature Collapse Strategy (AFCS), which incorporates an Anti-Collapse Domain Discriminator (ACDD) and the code rate loss function, effectively preventing excessive feature collapse. Furthermore, we propose a Flexible Feature Rebalance Module (FlexiReMod), a plug-and-play component that enhances generalization and dynamic feature capture through feature fusion and attention mechanisms. Experimental results indicate AC-CDCN achieved 87.14% (±5.60) and 71.77% (±12.92) accuracy on SEED and SEED-IV datasets, underscoring its significant generalization advantage.

The Social Intuitionist Model emphasizes the primary role of moral intuitions, rather than analytical moral reasoning, in forming moral judgments. In line with this model, moral emotions, rather than reasoning, are strongly associated with moral actions. This study explored the impact of emotionally evocative film excerpts, specifically those featuring moral dilemmas within varying contexts of social distance, on moral judgment processes. We invited 88 college students to view excerpts from foreign movies and their Chinese remakes under three music conditions: no music, positive music, and negative music. Participants evaluated the characters' actions based on three moral dimensions: Harm, Fairness, and Authority, while also rating their emotional responses. The results indicated that positive music during moral dilemmas elicited strong negative emotional responses. Additionally, viewers judged the protagonist's actions more harshly regarding fairness and placed greater emphasis on social order and authority in Chinese compared to foreign movie excerpts, regardless of the music's emotional valence.

Information disseminated through mass-media has been known to significantly influence behaviors such as voting (Iyengar & Kinder, 1987), brand preferences (Tversky & Kahneman, 1981), and public opinion (McCombs & Shaw, 1972). More specifically, the way information is framed and presented in mass media (for instance through "breaking news" and "sensational headlines") can reinforce existing beliefs and contribute to political and ideological polarization, with partisan media creating "echo chambers" that deepen biases (Hobolt et al., 2024). The current study leverages these insights to investigate the effects of the echo-chambered media and the time available for information consumption in an intergroup context across three experiments. Going by the minimal group paradigm, Experiment 1 employed a randomized, untimed presentation (non-echo-chamber) of news about an ingroup and an outgroup, while Experiment 2 used a blocked, untimed design (echo-chamber), and Experiment 3 a blocked, timed design (echo-chamber, doomscroll). All experiments involved two news sources with varying reliability (low and high) disseminating valanced (positive/negative) intergroup news headlines, and the participants were asked about the degree to which they believed the specific news items. Results showed that the manner of news presentation and timing moderated ingroup favoritism, with higher propensity to believe positive ingroup than positive outgroup news and vice-versa for the negative news, with the most significant bias emerging in Experiment 2 and negative news bias in Experiment 3. These findings shed light on how patterns of media consumption may influence intergroup perceptions and lead to polarization in society.

Modeling episodic memory (EM) remains a significant challenge in both neuroscience and AI, with existing models either lacking interpretability or struggling with practical applications. This paper proposes the Vision-Language Episodic Memory (VLEM) framework to address these challenges by integrating large-scale pretrained models with hippocampal attractor dynamics. VLEM leverages the strong semantic understanding of pretrained models to transform sensory input into semantic embeddings as the neocortex, while the hippocampus supports stable memory storage and retrieval through attractor dynamics. In addition, VLEM incorporates prefrontal working memory and the entorhinal gateway, allowing interaction between the neocortex and the hippocampus. To facilitate real-world applications, we introduce EpiGibson, a 3D simulation platform for generating episodic memory data. Experimental results demonstrate the VLEM framework's ability to efficiently learn high-level temporal representations from sensory input, showcasing its robustness, interpretability, and applicability in real-world scenarios.

Infants' attachment to their caregivers is a central feature of their early social and emotional development. Attachment Theory posits that these relationships vary systematically across distinct styles, though there has been debate about the extent to which these differences reflect features of caregivers' responsiveness vs. infants' own temperament. We develop a simple reinforcement learning model of infant exploration that allows us to vary the characteristics of simulated infants and caregivers and analyze the resulting patterns of model behavior. A set of equilibria reliably emerges that corresponds qualitatively to canonical attachment styles; particular agents' equilibria are controlled by both caregiver and infant parameters. These simulations point the way towards a quantitative synthesis of prior theoretical debates about the nature of attachment.

Sign languages are perceived visually and externalized using a signer's hands, face, and upper body. During sign language comprehension, deaf signers primarily focus their gaze on the face, while hearing non-signers attend more to the hands of a signer. Little is known about whether deaf signers adapt their gaze behaviour when processing unknown signs. Here, we report eye-tracking data from 15 deaf native signers of German Sign Language (DGS) and 15 hearing non-signers who were presented with videos in either DGS or an unknown sign language, all containing no linguistic mouth actions. Our data confirm that deaf signers generally fixate more on the face of a signer than hearing non-signers who attend to the hands in sign space. Moreover, only deaf signers increase their attention to the hands when processing video stimuli consisting of unknown signs compared to familiar signs, suggesting similar adjustment behaviours as observed in spoken languages.

Creativity is a complex and multifaceted construct that has been linked to various cognitive, emotional, and personality factors. This study explores the interplay between fluid intelligence, creative reasoning, personality traits, and mood states in predicting divergent thinking performance. We hypothesize that extraversion, openness to experience, positive moods, and creative reasoning would predict performance on divergent thinking tasks. A total of 120 young adults participated in the study, completing assessments on fluid intelligence (APM), creative reasoning (CRT-Reasoning), creativity (CRT-Creativity, TCT-DP), personality traits (Big Five Inventory 2), and moods (Current Mood Scale). Results revealed a negative relationship between fluid intelligence and divergent thinking performance, suggesting that higher fluid intelligence may be associated with more convergent thinking strategies. Extraversion emerged as a significant positive predictor of divergent thinking, supporting the idea that sociability and external engagement foster creativity. Openness to experience did not significantly predict divergent thinking, indicating that its impact may vary across domains of creativity. Mood states, especially hopelessness, were negatively associated with creativity, but the frequently found association between positive moods and creativity could not be replicated. These findings underscore the importance of considering cognitive, emotional, and personality factors in the study of creativity and suggest potential pathways for future research into how these elements interact to shape creative thinking.

Computer-based assessments enable the collection of fine-grained response process data through log files. We propose a novel method for extracting latent dimensions from such multidimensional response timing data, based on applying the Weighted MDS (WMDS) model. In our method, dissimilarities among examinees in their response timing vectors are computed, one such matrix for each test item, then WMDS is applied to this collection of matrices. The resulting latent dimensions represent variation among examinees in their patterns of response timing variables, with the dimension weights of the WMDS model reflecting differences across items in the importance of the latent dimensions. Latent dimensions are interpreted via permutation-based importance, correlation analysis and network analysis. Our method is demonstrated using response data from the PISA 2018 Reading and Mathematics assessments. Results show that the extracted latent dimensions are statistically reliable, educationally interpretable, and boost predictive accuracy when used in conjunction with item scores.

This research examines how language (generic vs. specific) and speaker knowledge (knowledgeable vs. unknowledgeable) influence essentialist beliefs about a novel social category in children and adults. Across two studies (N = 448 children, 433 adults), adults were more likely to endorse essentialist beliefs when knowledgeable speakers used generic descriptions. Children's responses varied by evaluation timing. In Study 1, when test questions were delayed, children's essentialist beliefs were influenced by language but not speaker knowledge. In Study 2, when memory demands were reduced by having children evaluate claims immediately after hearing them, children showed sensitivity to speaker knowledge, mirroring adults' responses. These findings highlight the role of language and contextual cues in shaping essentialist thought about social groups, suggesting that the effects of generics on social thought are dependent on the cultural expertise of the speaker.

Stereotypes change over time and across cultures, but they are hard to change in experiments. This paper proposes a solution by reconceptualizing stereotypes not as simple associations between groups and traits, but as rich networks of interconnected concepts spanning multiple domains. Building on cross-domain mapping, we used a simple question-answering task (''If X were a Y, what Y would it be?'') to create the expansive networks of 100 social groups to eight domains including animals, jobs, sports, colors, beverages, vehicles, musical instruments, and academic subjects. We found, for instance, the stereotype of women lacking agency is part of a larger network where women are associated with preferences for certain drinks such as wine, colors of pink, musical instruments of harps, or sports such as softball. We further tested whether rewiring these broader networks could change stereotypes more effectively than prior methods. Network-based interventions showed promising results for some groups (women were seen as more competent and Muslims were viewed as more friendly), but effects varied for different groups (minimal changes for criminals). This work suggests that successful stereotype change may require engaging with broader networks of subtle, seemingly unrelated associations rather than targeting individual stereotypical beliefs in isolation.

We propose a minimalistic representational model for the head direction (HD) system, a crucial component of spatial navigation in mammals. Our model leverages the symmetry of the rotation group U(1) and the inherent circular geometry of the head direction. We develop fully connected and convolutional versions of the model, both aiming to learn a high-dimensional representation of head direction that captures essential properties of HD cells. Our learning method results in representations that form a perfect 2D circle when projected onto their principal components, reflecting the underlying geometry of the problem. We also demonstrate that individual dimensions of our learned representation exhibit Gaussian-like tuning profiles akin to biological HD cells. Our model achieves accurate multi-step path integration, effectively updating its internal heading estimate over time. These results demonstrate that simple symmetry-based constraints can yield biologically realistic HD representations, offering new insights into the computational principles underlying spatial navigation in mammals.

Some people like coffee while others prefer tea, but little is known about whether preferences like these are shared among groups and whether they vary systematically across many common semantic categories. This study addresses this gap by examining two major sources of variation -- gender and extraversion -- across twelve categories or domains, ranging from fruit and animals to sports and personal qualities. In Study 1, participants rated their own preferences for a set of 300 exemplars. Results showed significant preference differences between men and women for 40% of items spread across all categories, and smaller but reliable differences between introverts and extraverts for 11% of items concentrated in domains like personal qualities. Study 2 used an allocentric categorisation task where the same participants categorised items based on which they thought would be preferred by men vs women or introverts vs extraverts. Using the ratings from Study 1 to score accuracy, the judgments from Study 2 showed that participants were sensitive to even subtle differences in preference, although accuracy varied by the judge's gender and extraversion: women were more accurate than men across many categories and introverts more accurate than extraverts for a few categories. We also found incorrect but widely shared judgments for about 20% of items, suggestive of inaccurate stereotypes about group preferences. Together, these results suggest widespread and systematic variation by gender (and to a lesser extent extraversion) that can be accurately predicted by others, although with systematic biases. Our results have implications for theories of semantic representation and social cognition.

Knowing the intended use of an artifact impairs people's ability to think of alternative uses. Here we ask whether children consider not just "what" a tool is used for, but also "who" uses the tool in that way. We focused on gender roles since children are sensitive to these early in development and adapted a classic insight learning task (Defeyter and German 2003). Success on the task requires inserting a stick into a tube to remove a ball. We compared children's (N = 112; 27 children/condition) latency to solve the problem at Baseline and three Demonstration conditions. In all the Demonstration conditions, the long stick was used as a magnet wand to brush away iron filings. In the Researcher Demonstration condition– a direct replication of Defeyter and German 2003– this function was demonstrated by a single individual – the experimenter; in the Same and Different Gender Peer Group conditions, the function was demonstrated by a group of children whose gender matched or differed from the participant's. Both Peer Group Demonstration conditions induced functional fixedness comparable to the Researcher Demonstration, and children were slower to solve the problem in all Demonstration conditions than Baseline. Critically however, children were faster to solve the problem in the Different Gender Peer Group condition than in the Same Gender Peer Group Condition, suggesting that children encode attributes of a function's typical user in their representations of artifacts, and that functional fixedness is affected by children's identification with a social role.

Emotional prosody---vocal cues that convey affect---profoundly influences how listeners interpret a speaker's intentions. We conducted two studies comparing AI- and human-generated emotional speech. In Study 1 (N=38), participants categorized five emotions (happy, sad, angry, neutral, fear) expressed by human voices and by an advanced text-to-speech (TTS) system. Human recordings exhibited higher overall accuracy (79.82% vs. 72.65%) and were rated significantly more natural, an effect partially explained by micro-perturbations (e.g., jitter, shimmer) that enhanced perceived authenticity. In Study 2 (N=53), these validated stimuli were incorporated into short scenarios, with each speaker labeled as either ``human'' or ``AI.'' Even when participants heard identical clips, those informed that the speaker was human exhibited greater trust and empathy, resulting in higher donation and advice-following rates. Although contemporary TTS systems effectively convey broad affective states, explicit AI labeling reduces perceived credibility and social engagement, underscoring the critical role that preexisting expectations play in human--AI communication.

This research investigates the feasibility of a communication system using Electroencephalography (EEG) signals mediated by a deep learning model, designed to aid individuals with significant communication challenges. The study builds on prior work in EEG-based image generation, testing the accuracy and efficiency of the system in conveying intended meanings. In the experiment, participants were presented images generated from EEG signals and were asked to give titles according to their interpretation. These titles were analyzed using text embeddings derived from a large language model (LLM) to measure cosine similarity. Results indicate that while sender and receiver interpretations often diverged, consistent patterns emerged within and between receivers. This suggests that repeated communication trials will align interpretations over time, improving mutual understanding. The findings highlight the potential of the method to facilitate adaptive communication, though further research is required to optimize its reliability, scalability, and practical applicability.

In contemporary consciousness studies, the sense of body ownership (SBO) stands as a key marker of subjective embodiment and self-awareness. Recent progress in multimodal and agent AI has prompted the question: Could an artificial system develop something analogous to SBO, and would this require consciousness? This paper refines two core distinctions: (1) functional versus phenomenal SBO, (2) world‑model versus self‑model. Building on a functional reconceptualization of "body" as an interaction boundary, we argue that AI systems equipped with semantic-centric multimodal world models and complementary self-models can, in principle, instantiate a form of SBO. By integrating diverse sensory inputs—visual, tactile, and linguistic—into a cohesive self-representation, this approach suggests the possibility of a virtual body that evokes the contours of the human embodied experience. Such an account questions the strict divide between physical and virtual embodiment, offering new insights into how embodied cognition underpins consciousness. Also, we locate SBO within the broader debate on AI consciousness. Concrete design proposals are linked to existing multimodal agents (e.g., DeepMind's MIA). This inquiry highlights how AI SBO may arise from the interplay of sensory and semantic frameworks, prompting ethical and theoretical reflection on AI consciousness.

Short-term phonetic accommodation is a fundamental driver behind accent change, but how does real-time input from another speaker's voice shape the speech planning representations of an interlocutor? We advance a computational model of change in speech planning representations during phonetic accommodation, grounded in dynamic neural field equations for movement planning and memory dynamics. A dual-layer planning/memory field predicts that convergence to a model talker on one trial can trigger divergence on subsequent trials, due to a delayed inhibitory effect in the more slowly evolving memory field. The model's predictions are compared with empirical patterns of accommodation from an experimental pilot study. We show that observed empirical phenomena may correspond to variation in the magnitude of inhibitory memory dynamics, which could reflect resistance to accommodation due to phonological and/or sociolinguistic pressures. We discuss the implications of these results for the relations between short-term phonetic accommodation and sound change.

Human experience is rich with sensory information, from which physical regularities are internalized to guide behavior. This paper presents the Patchwork Approach, a method for modeling sensorimotor predictions based on these regularities, without explicitly encoding physics. This method leverages environmental regularities, enabling resource-rational perceptual predictions. Using data from previous studies (Deeb et al., 2021, 2024) on human perception, the model outperforms Newtonian-based models in capturing human prediction errors and interpolates across untested conditions. One test demonstrates its superiority in a projectile motion task, while another illustrates its ability to predict deflection angles in a collision event from untrained aiming angles. This approach provides valuable insights into how physical laws are internalized and used to guide perception and action–arguing that perception provides foundational input to intuitive physics reasoning, a role that can complement and be extended to higher-level cognitive processes like those explained by the Intuitive Physics Engine (IPE).

A widely studied cognitive bias in decision-making is the attraction effect, in which introducing a clearly inferior decoy option into a binary choice set increases the likelihood of selecting the dominating option (target) over the competitor. While this effect is robust with numerical stimuli, findings with perceptual stimuli have been inconsistent, with some recent studies even reporting a negative attraction effect. We argue that this distinction between numerical and perceptual stimuli is superficial, and that choice behavior is better explained by the inter-attribute relationships. In two experiments, we demonstrated positive attraction effects using combined perceptual-numerical stimuli and both positive and null effects with numerical stimuli by manipulating the asymmetry in pairwise comparison difficulty. The numerical stimuli used in Experiment 2 leveraged the findings from fraction research. Together, our results challenge the numerical-perceptual distinction and support a universal cognitive mechanism underlying the attraction effect.

Interactions between agents are supported through a continuous process of detecting and responding to behaviors that are contingent upon the other agent's behavior. Here, we explore the temporal dependence of these mechanisms, focusing on the role of timescale compatibility in inter-agent interactions. Using continuous-time recurrent neural networks (CTRNNs) to control embodied agents in a minimal social interaction task, we demonstrate that effective interactions require agents to operate on compatible timescales. Our results indicate that time scale mismatches disrupt agents' ability to distinguish other agents from non-social entities, revealing a timescale threshold beyond which agents begin misclassifying slower agents as static objects and faster agents as non-social animate objects.

Real-world categories often exhibit graded structure, yet learners struggle to acquire family resemblance categories compared to unidimensional ones in laboratory studies. We propose that part of this difficulty arises from the binary nature of Traditional Artificial Classification Learning. We introduce Graded Classification Learning, a paradigm integrating category and quality judgments into response and feedback phases of a learning trial. This allows higher fidelity feature space exploration, aligning more with naturalistic learning processes. The ‘graded' learners showed superior performance, with higher final accuracy and steeper learning curves than the ‘traditional' learners. While aggregate response patterns appeared similar across conditions, profile analysis revealed an apparent gradedness in the traditional condition that was masked by an overwhelming preference for (bisecting) unidimensional strategies, whereas graded participants mostly exhibited genuine graded responses. These findings suggest traditional binary tasks may inadvertently hinder learning of graded structure and that incorporating quality judgments fosters robust category representations.

Stopping decisions are frequently modeled as decisions to switch to alternative activities once the current activity stops being adequately rewarding, such as in optimal foraging theory, as well as more recent metacognitive models. However, the sense of stopping and making decisions in such frameworks is highly platonic, with both decisions and stopping actions occurring instantaneously. In contrast, the phenomenology of quitting actions that one is undertaking appears to be temporally extended and metacognitively challenging. We study the metacognitive covariates of quitting decisions made by chess players using a large database of chess games sourced from an online chess portal. Our analysis reveals that players tend to persevere when they are playing against stronger opponents and after having played poor moves. We also find that a history of quitting games makes players more likely to quit in future games, but that having recently quit in a game offers some protective effect against quitting. Finally, we find that quitting a game makes it more likely that a player will play a game again soon. We place these results in the context of modeling quitting as a metacognitive choice affected by multiple competing goals.

The increasing integration of large language models (LLMs) based conversational agents into everyday life raises critical cognitive and social questions about their potential to influence human opinions. Although previous studies have shown that LLM-based agents can generate persuasive content, these typically involve controlled English-language settings. Addressing this, our preregistered study explored LLMs' persuasive capabilities in more ecological, unconstrained scenarios, examining both static (written paragraphs) and dynamic (conversations via Telegram) interaction types. Conducted entirely in Hebrew with 200 participants, the study assessed the persuasive effects of both LLM and human interlocutors on controversial civil policy topics. Results indicated that participants adopted LLM and human perspectives similarly, with significant opinion changes evident across all conditions, regardless of interlocutor type or interaction mode. Confidence levels increased significantly in most scenarios. These findings demonstrate LLM-based agents' robust persuasive capabilities across diverse sources and settings, highlighting their potential impact on shaping public opinions.

Learning motivation is a key variable in learning. Therefore, its assessment has consistently been a popular research topic. While traditional methods like self-report still dominate, methods integrating passive mobile sensing have emerged, using smartphones to collect behavioral data and assess learning motivation via statistical and machine learning techniques. Recent advances in large language models (LLMs) offer new perspectives for psychological measurements, yet their application in learning motivation assessment remains underexplored. To bridge this gap, we propose a novel approach that integrates LLMs with passive mobile sensing to assess and predict students' learning motivation. We constructed our dataset using mobile sensing data and self-report measures, then designed zero-shot and few-shot tasks embedded in LLMs to evaluate the performance. Ultimately, our findings indicate the feasibility and highlight the potential of leveraging LLMs to predict learning motivation levels based on mobile sensing data.

Detecting depression from user-generated posts on social media platforms offers significant potential for early intervention on at-risk individuals. Existing works mainly concentrate on text processing, and only a limited number incorporate images posted by users. These image-integrated methods face challenges in modeling the intricate relationships between textual and visual features. Besides, the absence of approaches that explore psychological trajectory of users by analyzing their posts over time leaves a critical gap in capturing the progression of depressive symptoms. In this paper, we propose A Time-Aware Mental State Space (T-M2S) for detecting depression from social media posts. We introduce a Cross-Modal Learning that effectively integrates text and image embeddings into sentiment-oriented unified representations. Additionally, we design a Mental State Space to analyze users' posts over time, offering a nuanced understanding of emotional dynamics. Extensive experiments on Twitter and Reddit datasets demonstrate that T-M2S significantly outperforms state-of-the-art methods. Code and models are available at GitHub.

Cross-Domain Sequential Recommendation (CDSR) predicts user behavior by leveraging historical interactions across multiple domains, focusing on modeling cross-domain preferences through intra- and inter-sequence item relationships. Inspired by human cognitive processes, we propose Hierarchical Attention Fusion of Visual and Textual Representations (HAF-VT), a novel approach integrating visual and textual data to enhance cognitive modeling. Using the frozen CLIP model, we generate image and text embeddings, enriching item representations with multimodal data. A hierarchical attention mechanism jointly learns single-domain and cross-domain preferences, mimicking human information integration. Evaluated on four e-commerce datasets, HAF-VT outperforms existing methods in capturing cross-domain user interests, bridging cognitive principles with computational models and highlighting the role of multimodal data in sequential decision-making.

In perceptual decision-making tasks, humans integrate noisy sensory evidence over time to guide their choices. The optimal integration process assumes that all evidence is weighted equally within a trial and that different trials are independent. However, humans exhibit systematic deviations from optimality, including uneven weighting of evidence within trials and influences from previous trials. Prior studies have demonstrated that biological constraints can account for this suboptimality. In this study, we present evidence that humans adapt their evidence integration strategies over time in response to task demands, and that the suboptimal uneven weighting is gradually learned over the course of the task. By explicitly modeling this adaptation through online gradient-based learning, our model outperforms existing approaches in capturing human behavior and unifies both observed forms of suboptimality in the Click task: dependence across trials emerges from an error-driven learning process that also gives rise to uneven integration weights within trials. We further propose a bounded-rational adaptation account to explain why humans progressively learn to weight evidence unevenly within a trial. Our modeling framework provides a general approach of resource-rational adaptation. It captures how initially uninformed agents can gradually update their strategies through error-driven learning and is applicable to a broad range of learning and decision-making scenarios.

Research has shown that infants prefer prosocial characters over antisocial ones, suggesting that sociomoral evaluation is early-emerging. However, some have argued that infants' preferential responses stem from low-level perceptual processes rather than true social understanding. Using electroencephalography (EEG), past work has suggested that motivational and social, but not attentional, processes are implicated in infants' responses to prosocial versus antisocial acts and individuals, however, the majority of past work utilized a single type of prosocial/antisocial interactions: helping a character to climb a hill. To test the generalizability of past neural findings from the hill paradigm, here we examined infants' responses in a distinct helping/hindering scenario in which a character tries but fails to open a box and is alternatively helped or hindered. Largely replicating past work, infants showed greater activity in social (indexed by the P400) but not attentional (indexed by the Nc) ERP components when seeing hinderers versus helpers, consistent with claims that infants' responses to prosocial and antisocial agents are social. No evidence of differential approach/avoidance motivation during prosocial/antisocial events was found. These findings support the role of social processes in infants' sociomoral evaluations.

People often rely more on their prior beliefs than the presented evidence when evaluating arguments. We investigate the cognitive mechanisms underlying this phenomenon. We hypothesise that when individuals encounter an argument that is congruent with their beliefs, it activates related information in memory. For belief-congruent arguments, people should therefore be more likely to both correctly recognise previously encountered information and incorrectly recognise new information as previously seen. To test this, we first investigated the effect of participants' beliefs about political claims on their evaluation of corresponding arguments that varied in quality. We then employed a surprise memory test to assess participants' recognition memory for these arguments. While we replicated the finding that prior beliefs drive argument evaluations, prior beliefs did not affect memory performance for all arguments in the same way. Our results indicate that individuals may use prior beliefs to aid memory only when the memory task is difficult.

Computational modelling offers a powerful tool for formalising psychological theories, making them more transparent, testable, and applicable in digital contexts. Yet, the question often remains: how should one computationally model a theory? We provide a demonstration of how formalisms taken from artificial intelligence can offer a fertile starting point. Specifically, we focus on the "need for competence", postulated as a key basic psychological need within Self-Determination Theory (SDT)—arguably the most influential framework for intrinsic motivation (IM) in psychology. Recent research has identified multiple distinct facets of competence in key SDT texts: effectance, skill use, task performance, and capacity growth. We draw on the computational IM literature in reinforcement learning to suggest that different existing formalisms may be appropriate for modelling these different facets. Using these formalisms, we reveal underlying preconditions that SDT fails to make explicit, demonstrating how computational models can improve our understanding of IM. More generally, our work can support a cycle of theory development by inspiring new computational models, which can then be tested empirically to refine the theory. Thus, we provide a foundation for advancing competence-related theory in SDT and motivational psychology more broadly.

Human language acquisition involves diverse learning resources, including self-supervised learning (sequence prediction) and communicative interactions (talking to caregivers). While recent advancements in language models highlight the power of self-supervised learning, the role of communicative interaction remains unclear. This study uses Reinforcement Learning (RL) and parent-child agent simulations to model interactions and investigate their role in language acquisition, as well as whether RL-like mechanisms may function in children. We pretrained a small transformer model as a child agent, which then interacted with Google's Gemini, acting as a parent agent, to learn language with the goal of being understood. Model evaluations show that the interactive training enhances intelligibility of model's communication and increases behavioral similarity to real child speech. However, minimal pertaining alone provides noticeable syntactic and semantic competence, with RL yielding no consistent gains. These findings imply that interaction may play a more critical role in pragmatic aspects of language learning than in the development of linguistic structures, and that learning through interaction is a mechanism used by children.

Children with Developmental Language Disorder (DLD) exhibit longer reaction times (RTs) than age-matched neurotypical children. Drift diffusion models estimate parameters influencing RT distributions: drift-rate represents speed of information accumulation, non-decision time represents other factors contributing to longer RTs, e.g., poor attention or motor coordination. Using a hierarchical Bayesian framework, we modeled RT data from visual search and mental rotation tasks completed by 3rd graders (N = 248). Children with impaired verbal abilities without accompanying nonverbal impairment (DLD) and those with global verbal/nonverbal impairments were compared to neurotypical children. Across tasks, children with DLD exhibited a lower drift rate than neurotypical children, indicating slower information accumulation, and no difference in non-decision time. Children with global impairments showed lower drift rates and higher non-decision times than neurotypical children. Results suggest directions for nuanced tests of the generalized slowing hypothesis of DLD. Clinical implications for diagnosis and treatment of DLD are discussed.

Being bilingual is becoming increasingly common for children worldwide, including those with autism spectrum disorder (ASD). While recent studies have examined the effect of bilingual experience on cognitive and linguistic abilities in autistic children, few have focused on how autistic status influences bilingual language switching and control. For the first time, the present study investigates language control during bilingual word production in autistic and typically developing children with a cued language switching paradigm. Results revealed that autistic children tended to make more cross-language mistakes and had more difficulties when switching between languages, while the overall naming latency and language mixing costs were similar across the two groups. The preliminary findings highlight the potential challenges encountered by autistic children on different levels of language control during bilingual production and also suggest that some aspects of language switching performance are comparable between the groups. Clinical implications are also discussed.

Emotion regulation is a fundamental process that shapes cognitive, affective, and behavioral responses to emotional stimuli. Traditional emotion regulation models conceptualize regulation as a sequential modulation of emotional responses. However, they do not fully explain how emotions are constructed in a way that allows such regulation to occur. Predictive processing (PP) provides a mechanistic framework for understanding emotion generation by proposing that the brain minimizes prediction errors (PEs) to optimize perception and behavior. Yet, standard PP accounts reduce valence to PE minimization, failing to explain how PEs can generate different subjective feelings. To address these limitations, we propose a valence-focused model of emotion regulation that integrates predictive processing with self-referential cognition. We incorporate emotional valence as interpretative processes of the self-model, which assigns emotional significance based on goals, values, and autobiographical context. This model bridges the gap between emotion generation and regulation, highlighting the dynamic interplay between prediction errors, subjective valuation, and self-referential processes. This approach not only advances theoretical understanding but also opens new avenues for computational modeling and empirical research into the adaptive and maladaptive aspects of emotional experience.

Affective Brain-Computer Interfaces has achieved remarkable advancements, enabling researchers to interpret labeled EEG data accurately. However, the annotation of EEG data is time-consuming and requires substantial effort, which limits the application in practical scenarios. In this paper, we propose a self-supervised EEG representation learning framework based on temporal prediction and spatial reconstruction (EEG-TPSR) to learn EEG representations from a large amount of unlabeled data. Our model consists of two stages: 1) In the pre-training stage, we use contrastive temporal prediction and spatial reconstruction as proxy tasks, which utilize the spatio-temporal information to learn the generic representations from EEG data; 2) In the fine-tuning stage, few data is used to calibrate the pre-trained model. We conduct extensive experiments on three emotion EEG datasets. The results demonstrate that our proposed model achieves excellent performance, with over 20% relative accuracy improvement and more than 15% improvement using only 1% labeled data.

Children's emotional development is linked to language development for typically developing children and deaf children with native sign language exposure. However, approximately 90% of deaf children are born to hearing parents who are not familiar with sign language. These deaf children begin learning a sign language when they attend a school for the deaf. Late sign language exposure has negative consequences on several aspects of language development. We investigate whether acquiring sign language late affects children's emotion recall and channel of emotion expression. After watching a silent video depicting emotions, late- and native-signing deaf children retold the story in Turkish Sign Language. Results showed that late signers recalled fewer emotions and used fewer signs and facial expressions compared to native signers. Manual gestures (non-sign hand movements), head and body movements did not differ across groups. The findings suggest that late sign language acquisition negatively impacts deaf children's ability to recall and express emotions, highlighting the importance of early language exposure for the development of emotion recall.

Creativity assessment in science and engineering is increasingly based on both human and AI judgment, but the cognitive processes and biases behind these evaluations remain poorly understood. We conducted two experiments examining how including example solutions with ratings impact creativity evaluation, using a finegrained annotation protocol where raters were tasked with explaining their originality scores and rating for the facets of remoteness (whether the response is ``far'' from everyday ideas), uncommonness (whether the response is rare), and cleverness. In Study 1, we analyzed creativity ratings from 72 experts with formal science or engineering training, comparing those who received example solutions with ratings (example) to those who did not (no example). Computational text analysis revealed that, compared to experts with examples, no-example experts used more comparative language (e.g., ``better/worse'') and emphasized solution uncommonness, suggesting they may have relied more on memory retrieval for comparisons. In Study 2, parallel analyses with state-of-the-art LLMs revealed that models prioritized uncommonness and remoteness of ideas when rating originality, suggesting an evaluative process rooted around the semantic similarity of ideas. In the example condition, while LLM accuracy in predicting the true originality scores improved, the correlations of remoteness, uncommonness, and cleverness with originality also increased substantially --- to upwards of $0.99$ --- suggesting a homogenization in the LLMs evaluation of the individual facets. These findings highlight important implications for how humans and AI reason about creativity and suggest diverging preferences for what different populations prioritize when rating.

We approach the problem of explaining segmentation --- the human capacity to partition input streams into representations of appropriate form and content for efficient downstream processing --- by exploring a theoretically minimalistic and computationally plausible account of phoneme-to-word chunking. Through computational models, mathematical proofs, algorithm design, and observer model simulations in two languages, we suggest that online segmentation can be guided by content-agnostic properties of internal memory structures (i.e., lexicality and length type frequency). Our theoretical and empirical findings point to a formal link between such properties with practical performance benefits. Together, these contributions make progress on a fully explicit computational- and algorithmic-level account with plausible implementational-level primitives.

While research has explored cross-cultural variation in child-directed speech (CDS), little is known about if and how it may have changed over time. We explore whether CDS has undergone historical change by analyzing prosodic features in child-directed (CD) broadcasts from a German children's bedtime program (1959–present) and comparing them to adult-directed (AD) weather forecasts from the same period. The program originated in East Germany and continued after German reunification in 1990, potentially reflecting a socio-cultural shift toward more child-centric attitudes characteristic of Western liberal democracies. Pitch variation in CD broadcasts, although higher than in AD broadcasts, remained stable over time. In contrast, articulation rates showed no register difference pre-1990; only after 1990 did CD broadcasts exhibit the slower articulation rates typical of CDS. This suggests that some features of CDS may be subject to cultural evolution over historical time, which can be accelerated by major historical events.

Electroencephalogram (EEG) has become an important indicator reflecting emotions. Due to its natural graph structure characteristics, it has made significant progress in the emotional recognition using graph convolutional networks (GCN). However, existing methods face limitations: (1) insufficient integration of psychological prior knowledge, limiting the utilization of brain activity patterns, and (2) simplistic node relationship construction, neglecting the universality and functional connectivity of brain regions. Therefore, we propose a dual-path parallel graph convolutional network (DP-GCN). The first path leverages psychological prior knowledge to segment electrodes into brain regions and employs an attention mechanism to integrate features. The second approach employs a data-driven method, using a sparse stacked autoencoder to reconstruct brain region features, while a learnable, input-independent adjacency matrix captures EEG patterns associated with emotions. Finally, a cross-attention mechanism integrates features from both paths. DP-GCN has been evaluated on public dataset, achieving an accuracy of 82.69%±4.16%, demonstrating its competitive performance.

Previous studies of predictive mechanisms in computer-mediated communication (CMC) suggested that native speakers (L1) rely on auditory cues and emotion in conversation processing. To understand how the prediction mechanism differs for non-native speakers (L2) in CMC, this study assessed how the loss of multi-modal cues affects word predictability in turn-taking, considering language background and social factors. L2 watched videos, listened to audio, or read a transcript of conversations, and predicted the same set of omitted words with different levels of predictability and semantic relatedness in different CMC. Results showed that, similar to the L1 study in He et al. (2025), higher response similarity but longer response time were observed in conditions with richer cues in L2 predictive processing. Semantic relatedness, self-emotion, attention, and language proficiency did not affect predictability. Participants reporting negative emotions and more limited L2 exposure demonstrated reduced prediction accuracy, particularly in cue-rich environments. These findings expand our understanding of L2 predictive processing in CMC by highlighting how multimodal cue integration operates differently for L2. The results have implications for developing communication technologies and language pedagogies tailored to L2 across various mediated communication contexts.

To advance our understanding of referential communication and common ground formation, this study presents a novel generative cognitive model that integrates deep neural networks for visual perception, image generation, and language captioning. Using the Tangram Naming Task (TNT), we simulate the sender–receiver interaction with modular processes replicating holistic cognitive strategies. Through controlled simulation experiments, we reveal that language generation plays a more crucial role than visual perception in establishing common ground, while intermediate image generation enhances linguistic diversity—a key aspect of natural communication. Our results bridge cognitive modeling and large generative models, demonstrating how internal cognitive dynamics can be visualized and quantitatively evaluated. This study contributes to the growing field of cognitive-inspired human–AI communication and provides a blueprint for grounding-rich simulations in collaborative tasks.

A growing body of research suggests the presence of an attentional component in the memory advantage of animate items. Most research on this animacy effect is focused on remembering, but its effects on forgetting are less well-researched. With the use of pupillometry, we investigated the attentional processes present in the selective remembering and directed forgetting of animate and inanimate items. More specifically, we investigated whether external and internal attention are affected by the animacy status of to-be-remembered and to-be-forgotten words in an item-method directed forgetting task with retro cues, followed by a recognition task for all previously presented items. Our behavioral results demonstrate the directed forgetting effect: accuracy on the recognition task was higher for to-be-remembered words than for to-be-forgotten words. Additionally, we found an advantage of animacy, with higher accuracy for animate words than for inanimate words, regardless of cue. The pupillometry results demonstrate differential internal attention for to-be-remembered and to-be-forgotten words during cueing, with larger pupil sizes for remember cues than for forget cues, but no effect of animacy. We did however find an animacy effect on external attention during stimulus presentation, with smaller pupil sizes for animate words than for inanimate words, which may reflect precision of encoding. This suggests that not internal attention, but external attention is influenced by animacy status, potentially explaining the memory advantage of animate items through the richness-of-encoding hypothesis.

Systems of semantic categories vary across languages, but this variation appears to be constrained by pressure for optimizing a complexity-accuracy tradeoff known as the Information Bottleneck (IB) principle. This finding, however, has been based primarily on individual languages and it remains largely unknown how bilinguals navigate the category systems of two different languages, particularly when these languages' category boundaries do not overlap. Here, we address this gap in the literature by combining theory-driven experiments with an extension of the IB framework to bilinguals. Specifically, we investigate bilingual vs. monolingual category boundaries in English and Mandarin via a two-alternative forced-choice (2AFC) labeling task on six continua that interpolate between two distinct everyday objects (e.g., plate and bowl). We find that: (1) bilinguals do not maintain two monolingual-like systems but rather exhibit a converged semantic system influenced equally by both languages; and (2) this departure from monolinguals is nonetheless constrained by the same pressure for efficiency that operates in monolinguals. These findings provide new insight into how bilinguals navigate cross-linguistic semantic variation and suggest that despite having to accommodate myriad sociolinguistic factors, a drive for efficiency is also a key factor that shapes bilingual category systems.

Understanding human thoughts is a key goal of cognitive science. Behavioral observations alone limit insight into cognition. The think-aloud protocol, where participants verbalize thoughts, offers a direct probe into reasoning but is underutilized due to challenges in subjectivity and scalability. Advancements in natural language processing (NLP) enable computational analysis of think-aloud data, yet little work explores its role in strategy learning. We test whether think-aloud reports reveal strategy use in a stochastic learning task where participants verbalized their strategies. Our results show diverse strategy usage, with a preference for persistent choices. Think-aloud analysis suggests participants rely on distinct meta-strategies to guide learning. Clustering and predictive modeling reveal strong alignment between choices and verbalized strategies. These findings highlight think-aloud as a scalable tool with NLP techniques for studying high-level cognition, shedding light on a promising paradigm for cognitive sciences.

Mandarin Chinese is a tonal language where variations in voice pitch distinguish word meaning. Acquiring tonal contrasts presents challenges for adult second-language learners. Undergraduates (N = 59) completed a computer-assisted language learning protocol, where they engaged in listening and repeating Chinese disyllabic nouns and matching them with corresponding pictures. Tone production accuracy was assessed at pretest/posttest using complexity invariant distance (CID), a quantitative metric of the distance between time series (learner vs. native-speaker productions). Word-level analyses found lower CID scores at posttest, indicating improvements in tone production after three blocks of word-picture matching. Fine-grained syllable-level analyses showed lower CID scores for first vs. second syllables, suggesting a primacy advantage. Accuracy on the Music Ear Test predicted lower CID scores, linking musicality with aptitude in learning tonal contrasts. No effects of nonverbal intelligence or language background were found. CID offers a robust method of assessing tone production accuracy for future studies.

We frequently interact with others daily and experience a sense of joint agency—the feeling of performing an action together. Recent studies suggest that this sense of joint agency is influenced by the perceived "human-likeness" of partner. This study examined how a partner's behavioral process, specifically adaptation and fluctuation, affects joint agency in a cooperative task mediated by human-likeness. Participants completed a cursor-tracing task simulating collaboration, with cursor movement determined by combining their input with pre-recorded data. In this experiment, adaptation was approximated by preprogrammed changes in the cursor movement. The results revealed that adaptation enhanced joint agency, whereas fluctuation had no significant effect. Human-likeness is thus positively correlated with joint agency. Moreover, individual traits such as extraversion and attachment shaped these perceptions in unexpected ways. Poor task performance increases joint agency. These findings contribute to this field by identifying factors that influence the sense of joint agency.

Xenophobia and anti-immigrant sentiments have been increasing in Western democratic countries, and it is important to understand how messaging can improve attitudes towards immigrants. Past studies show prior attitudes are associated with how individuals evaluate related arguments. The present study (N = 349) explores if people's prior attitudes influence how they evaluate the strength of arguments in the context of immigration. We also test whether the style of argument (i.e., narrative or statistical) influences argument evaluation. We measured participants' attitudes towards immigrants before and after an argument evaluation task, where participants rated the quality of a narrative and statistical argument. Participants with high pre-existing negative attitudes towards immigrants rated pro-immigrant arguments poorly and anti-immigrant arguments strongly, and we see the opposite relationship for participants with pre-existing positive attitudes towards immigrants. Our findings demonstrate that people can evaluate the same arguments about immigrants very differently depending on their pre-existing attitudes and that argument style can affect argument evaluation.

The Speech-to-Song Illusion (STS) phenomenon, where repeated short speech utterances transform into perceived song, has drawn attention to its underlying mechanisms and cross-linguistic differences. This study examines the STS effects among Korean speakers, comparing standard Korean (non-tonal) and dialects such as Gyeongsang (pitch-accent, tonal) and Jeju (non-tonal but intonation-rich), which exhibit varying levels of linguistic tonal features. Participants (N = 60), evenly divided between standard and dialect users, evaluated 180 auditory stimuli comprising standard Korean, Gyeongsang, and Jeju utterances under controlled repetition conditions. Results revealed significant STS effects across all groups and stimuli, with stronger effects observed for dialectal stimuli, particularly Jeju, compared to standard Korean. Interestingly, differences between standard and dialect speaker groups in STS perception were not statistically significant, suggesting that exposure to diverse linguistic environments, facilitated by modern Korean media, may homogenize perceptual responses to tonal variations. The study highlights the influence of tonal and rhythmic elements in STS perception and underscores the cultural and linguistic uniqueness of Korean as a fertile ground for exploring auditory illusions. This research contributes to understanding the interplay of linguistic and perceptual factors in STS and opens avenues for cross-cultural comparisons and neuroscientific investigations of auditory illusions.

Visual working memory (VWM) plays a fundamental role in cognitive processes, such as perception, attention, and reasoning. However, existing approaches to modelling VWM are not integrated into cognitive architectures and lack interpretability with respect to their parameters. To address this limitation, we propose a novel VWM model based on the well-established Semantic Pointer Architecture (SPA). In contrast to previous works, our model is the first to integrate a VWM model with a cognitive attention model. It only requires three interpretable hyper-parameters: spatial capacity, feature certainty, and memory decay. We experimentally show that our base model without memory decay replicates the set-size effect and swap errors of human data on a continuous reproduction task. More importantly, we show that by introducing a memory decay, we can achieve a statistically significant (p ≪ 0.001) improvement in model fit, suggesting a potentially important role of memory decay in VWM. Further, our VWM model can be easily extended to model pre- and post-cue conditions, consistently achieving KL divergence between modelled and human performance of less than 0.05.

Communication is essential for enhancing the cognition and cooperation of agents in multi-agent reinforcement learning (MARL). However, existing methods often rely on predefined and rigid cognitive patterns, which cannot adapt to dynamic environmental changes and complex inter-agent interactions. In this work, we introduce the Hierarchical Cognitive Graph Autoencoder (HCGA), an adaptive framework that addresses these limitations. HCGA represents inter-agent messages as nodes in a graph with learnable edges, employs a grouping mechanism to integrate related local information into compact latent representations, and then applies hierarchical aggregation to construct a comprehensive global cognition. This approach effectively distills essential information and adaptively uncovers cognitive patterns from dynamic environments, thereby enhancing the overall robustness and efficiency of cognitive processing in MARL tasks. Experimental results demonstrate that HCGA significantly outperforms state-of-the-art methods across various MARL tasks, highlighting its robustness, adaptability, and efficiency.

Modern machine learning systems have demonstrated substantial abilities with methods that either embrace or ignore human-provided knowledge, but combining benefits of both styles remains a challenge. One particular challenge involves designing learning systems that exhibit built-in responses to specific abstract stimulus patterns, yet are still plastic enough to be agnostic about the modality and exact form of their inputs. In this paper, we investigate what we call The Ungrounded Alignment Problem, which asks How can we build in predefined knowledge in a system where we don't know how a given stimulus will be grounded? This paper examines a simplified version of the general problem, where an unsupervised learner is presented with a sequence of images for the characters in a text corpus, and this learner is later evaluated on its ability to recognize specific (possibly rare) sequential patterns. Importantly, the learner is given no labels during learning or evaluation, but must map images from an unknown font or permutation to its correct class label. That is, at no point is our learner given labeled images, where an image vector is explicitly associated with a class label. Despite ample work in unsupervised and self-supervised loss functions, all current methods require a labeled fine-tuning phase to map the learned representations to correct classes. Finding this mapping in the absence of labels may seem a fool's errand, but our main result resolves this seeming paradox. We show that leveraging only letter bigram frequencies is sufficient for an unsupervised learner both to reliably associate images to class labels and to reliably identify trigger words in the sequence of inputs. More generally, this method suggests an approach for encoding specific desired innate behaviour in modality-agnostic models.

Intrusive memories are disruptive to daily functioning and detrimental to well-being; unfortunately, the presence of these memories is a defining characteristic of post-traumatic stress disorder (PTSD). Although increasingly understood as a memory disorder, current trauma management and recovery strategies for PTSD do not often take principal memory theories into account. Many common practices, such as delayed processing after trauma exposure and spaced therapy sessions, might inadvertently strengthen the retention of intrusive memories in the long-term. In this paper, model simulations show that altering the timing of different presentations of emotional stimuli might affect subsequent intrusions. Experimentally, we demonstrated through a two-day within-subject image presentation task that when emotional images are presented in spaced intervals (as opposed to consecutive, "massed" presentation), the perceived frequency of intrusions for the mass-presented emotional images during the 24 hours after first exposure were significantly lower than spaced images. Our study presents a novel strategy that can potentially mitigate the frequency of intrusive post-traumatic memories, highlighting the advantages of translational applications of computational cognitive models to mental health.

Theories of music perception argue over whether observed motor area activation during passive music listening actively contributes to perception or is the product of a distributed representation. There is a growing amount of evidence linking Alpha rhythms in the motor cortex to action inhibition and imagination during passive music listening. In this work, we examine Alpha band power modulation and its association to beat perception using a sensorimotor synchronization task and a natural music listening task with electroencephalography (EEG). We sought to find an association between Alpha band modulation over the primary motor cortex and beat tapping performance. We found that greater Alpha power correlated with worse tapping performance. These results may point to a negative association between motor inhibition and beat perception and a complementary positive association between movement imagination and beat perception and production. We address these findings in terms of the HAPEM theory proposed by Schubotz (2003). This framework suggests that motor activation reflects a predictive representation formed from audio-motor association cortices, lacking proprioceptive information which could be acquired through musical training.

Understanding conditional inferences is fundamental to human reasoning, allowing us to predict the consequences of actions. For instance, the conditional, "If you eat your broccoli, you'll get a candy" can be interpreted literally, meaning eating broccoli is one way to get a reward, or pragmatically, implying it is the only way. Past studies show school-aged children (ages 7-12) struggle to arrive at literal meanings but, interestingly, compute adult-like, pragmatic interpretations at this age. A key limitation of past research is the lack of testing in contexts that favor literal meanings. We conducted two studies to examine whether children can derive literal interpretations when supported by context, focusing on scenarios where adults prefer literal over pragmatic interpretations. We found that preschoolers, as young as 4 years old, are adult-like in computing literal meanings of conditionals when contextually supported, and also can arrive at pragmatic meanings of conditionals. These findings inform theories of logical reasoning and implicature acquisition.

When we give feedback, we face a delicate balancing act – we want to convey accurate information, but we also don't want to hurt someone's feelings. While computational pragmatic models have elegantly shown how politeness emerges from these principles, they've mainly focused on choices from limited predefined responses. Large language models (LLMs) enable the study of open-ended politeness strategies, but their ability to balance informational and social goals like humans remains uncertain. First, replicate previous work using restricted utterance sets, finding that sufficiently large LLMs (≥70B parameters) capture key human politeness patterns, particularly the strategic use of negation. We then extend this investigation to open-ended contexts, collecting and evaluating naturalistic feedback from both humans and LLMs. Surprisingly, human evaluators preferred LLM responses, which demonstrated sophisticated goal sensitivity and diverse politeness tactics. These findings suggest remarkable pragmatic competence in LLMs' polite language generation while raising questions about the underlying mechanisms.

The time-varying appeal of an audiovisual stimulus cannot be directly observed because it is not only determined by the expression itself but also involves the viewer's information processing. In this study, we attempted to reconstruct a time-varying common input that is an appeal to induce blink rate synchrony, i.e., blink rate is suppressed at appealing scenes. In the experiment, 44 (22 male and 22 female) university students watched two videos promoting a local area in Japan while detecting blinks using an eye-tracking device. The results showed that the reconstruction ability was less dependent on the embedding parameters. The results showed that the peak of the reconstructed common inputs did not always correspond to the most impressive scene. In the future, it would be beneficial to apply this method to physiological index data according to the type of attractiveness.

This paper presents a categorization experiment supporting the hypothesis that base-rate-neglect occurs in linguistic category learning (and, thus, is a cross-domain phenomenon), and that it is more likely for those learners who engage in explicit problem-solving, rather than implicit learning. We find that among those participants who were able to verbalize the cue that was probabilistically associated with category-membership (correct-staters) in a phonological learning task, about a third respond in a way consistent with base-rate-neglect. On the other hand, non-staters are more likely to respond randomly or by probability matching the base-rates. These results suggest that explicit learning is associated with base-rate-neglect to a greater extent. We found that no learners integrate the two probabilistic patterns present in the experiment into a single Bayesian estimate. Rather, some of them focus on the base-rates, others focus on category-internal cues, and others simply fail to learn anything.

What drives an agent to explore the world while also maintaining control over the environment? From a child at play to scientists in the lab, intelligent agents must balance curiosity (the drive to seek knowledge) with competence (the drive to master and control the environment). Bridging cognitive theories of intrinsic motivation with reinforcement learning, we ask how evolving internal representations mediate the trade-off between curiosity (novelty or information gain) and competence (empowerment). We compare two model-based agents using handcrafted state abstractions (Tabular) or learning an internal world model (Dreamer). The Tabular agent shows curiosity and competence guide exploration in distinct patterns, while prioritizing both improves exploration. The Dreamer agent reveals a two-way interaction between exploration and representation learning, mirroring the developmental co-evolution of curiosity and competence. Our findings formalize adaptive exploration as a balance between pursuing the unknown and the controllable, offering insights for cognitive theories and efficient reinforcement learning.

Research on implicit learning using the artificial grammar learning (AGL) paradigm has traditionally relied on tasks that promote active engagement, such as memorization, repetition, or rule discovery during the exposure phase. This study examined whether limited exposure, devoid of active engagement tasks, enables participants to distinguish between grammatical and ungrammatical sequences in both simple and complex artificial grammars. Participants performed above chance on the grammaticality task across both conditions but appeared to rely on explicit strategies to a greater degree than reported in previous AGL studies. These findings highlight the critical role of exposure conditions and suggest that exposure to letter strings without active engagement may not sufficiently restrict learning to implicit processes.

Questioning has become increasingly crucial for both humans and artificial intelligence, yet there remains limited research comprehensively assessing question quality. In response, this study defines good questions and presents a systematic evaluation framework. We propose two key evaluation dimensions: appropriateness (sociolinguistic competence in context) and effectiveness (strategic competence in goal achievement). Based on these foundational dimensions, a rubric-based scoring system was developed. By incorporating dynamic contextual variables, our evaluation framework achieves structure and flexibility through semi-adaptive criteria. The methodology was validated using the CAUS and SQUARE datasets, demonstrating the ability of the framework to access both well-formed and problematic questions while adapting to varied contexts. As we establish a flexible and comprehensive framework for question evaluation, this study takes a significant step toward integrating questioning behavior with structured analytical methods grounded in the intrinsic nature of questioning.

The gradual nature of historical language change is widely acknowledged. We explore a syntactic change model that offers a new view into the theoretical difference between classical and neural network claims about language encoding. Most prior treatments of grammaticalization fail to account for how, exactly, new forms arise, focusing instead on change following innovation. A phenomenon relevant to the innovation puzzle is Quantitative Anticipation of Qualitative Change in Grammaticalization (QAQCG)---gradual statistical changes anticipate structural changes. Although prior researchers have given phenomenological descriptions, we know of no rigorous method for testing whether QAQCG exists. Here, we quantitatively examine the case of English "a lot" which has grammaticalized an Adverb function from a Noun Phrase function. A simple feedforward neural network implements QAQCG, predicting a curving trajectory in probability space. Bayes Factor analysis supports the network over a classically-motivated linear model, highlighting continuity and nonlinearity as distinctive theoretical claims.

Category labels such as `dog' and `green' appear to induce more categorical representations--highlighting category diagnostic features and helping to distinguish category members from non-members. Here we investigate whether covert language use has a similar effect by taking advantage of natural variation in people's reported use of inner speech. To measure categoricality, we use a repetition-priming task in which people make a semantic judgment of repeated images. We find a robust repetition effect of categories such that people are faster to respond to a cat if they have seen a previous image of a cat. These differences in inner speech are not associated with differences in repetition priming, but interacted in complex ways with differences in visual imagery and susceptibility to a verbal interference task.

People often have a preference for "self-verifying" feedback that confirms their existing self-views. Self-verification can reinforce existing self-views and prevent opportunities to learn from alternative perspectives, as when people with low self- esteem prefer feedback that validates negative self-beliefs. Past work suggests that a major driver of self-verification is a desire for accurate self-assessment, where disconfirmatory feedback that contradicts existing self-views creates doubt about the credibility of the feedback source. The aim of this study was to develop a formal account of self-verification based on a Bayesian model of source reliability. Findings from a behavioral experiment aligned with the model's prediction that confirmatory feedback about traits central to one's self-concept enhances the perceived reliability of a source, while disconfirmatory feedback leads to lower reliability and disinterest in further feedback. This approach clarifies why seemingly biased feedback seeking behaviors may be motivated by rational epistemic concerns about source credibility.

In the present study, we tested whether inducing people to reflect on their knowledge may increase their intellectual humility. We hypothesized that asking participants to answer knowledge tests would prompt them to recalibrate their perception of their own knowledge, thereby fostering intellectual humility. Study 1 demonstrated a significant increase in intellectual humility following the intervention, whereas Study 2 replicated and extended these findings in a larger sample, confirming the effect despite its small magnitude. The observed increase may be due to the activation of analytical reasoning style or to the acknowledgement of one's knowledge limitations. However, further research is needed to corroborate these conjectures and explore the long-term effects of interventions to enhance intellectual humility.

This study examined how classifier-noun conventions and classifier semantic compatibility influence the selection of classifiers in Mandarin- speaking children aged five to seven. Results indicated that children's classifier use was shaped not only by conventional associations between classifiers and noun categories but also by semantic congruence, particularly in the absence of explicit noun labels. Older children demonstrated greater sensitivity to labels in guiding classifier selection than younger children. Furthermore, explicit noun labels most strongly boosted children's choice of conventional classifiers for non‑prototypical stimuli, while having a much smaller impact on prototypical stimuli. Overall, these results highlight the interplay between memorization and semantic compatibility in classifier acquisition and underscore the importance of semantic and perceptual features in shaping language learning.

Emotions play a crucial role in social interactions, yet little is known about the effects of experiencing emotions together with others on expressed emotional valence and arousal. We compared changes in adults' body posture when recalling experiences of positive and negative basic emotions (happiness & sadness) and social emotions (pride & shame), either jointly (dyadic condition) or by themselves (individual condition). To capture the dynamic unfolding of the emotional experience, we used a novel depth sensor imaging technique based on LiDAR- technology integrated in a commercial tablet. Adults (N = 80) displayed greater postural chest-height elevation and upper-body chest expansion (measuring valence) following positive compared to negative emotion recalls. Furthermore, participants showed more overall movement (measuring arousal) after positive compared to negative emotions, especially in the dyadic condition. These results suggest that recalling emotions together affects non-verbal expression of emotions, and we discuss our findings in light of recent advances in emotion science.

Do simple spatial terms such as in or on map directly to intuitive physical judgements about spatial relationships between objects that underlie these terms' meaning? We explored this question in the domain of physical support. Adults (N=120) and 4-year-old children (N=42) were shown videos in which a puppet placed an L-shaped object in contact with a table at locations that varied in whether the object was supported or not. Half of the participants were asked for linguistic judgments ("Is X on Y?") and half were asked for intuitive physics judgments ("Will X fall if (agent) lets go?"). Results revealed that linguistic judgments were largely categorical, with child and adult participants labeling objects as on even when the object was not truly supported. In contrast, intuitive physics judgments aligned closely with the object's actual possibility of true support. However, responses also varied by the orientation of the L-shaped object, with on applying categorically to a regularly oriented L, but in a more graded fashion for a mirror image oriented L. Our findings suggest that the mapping between the simple spatial term on and physical reasoning systems are not completely coupled, and that the ways in which language draws on intuitive physical reasoning is complex.

Understanding the cognitive mechanisms of loan borrowers at peer-to-peer (P2P) loan platforms is helpful for improving communication strategies during loan marketing. Recent research studies the cognition of loans based on field studies or interviews with limited research samples, which cannot gain comprehensive cognitive insights in real-world context. In this work, we use a concept mapping-driven method and a large amount of real-world data for loan cognitive analysis. We find that there are statistical differences between the concept mappings of loan borrowers who received lenders' support and those who did not. We also found the representative conceptual factors of borrowers that impact the lenders' decision-making. For example, lenders are more inclined to support borrowers who present a mindset focused on stability, adaptability, and purposeful transformation, whereas non-recipients often express uncertainty, lack of readiness, or superficial changes in their loan requests. Applicants with an abstract cognitive orientation tend to request loans with higher interest rates, in contrast to those with a more concrete conceptualization. In practice, loan borrowers and lenders can refer to the cognitive findings to support their marketing and decision-making.

Understanding and operating on multi-digit numbers is a critical step in the development of mathematical skills and concepts. Empirical and computational modeling evidence suggests that multi-digit numbers are processed decomposed into units, tens, hundreds, etc. for magnitude comparison. Yet, there is currently no computational model to simulate multi-digit number arithmetic. Accordingly, we developed a neural network model with three interconnected modules reflecting single-digit additions, recognition of a carry-over, and decision-making. We then compared model performance after following two training curricula: i) a \textit{ step-by-step} curriculum, where single-digit addition are trained before progressing to multi-digit problems and ii) an \textit{all-at-once} curriculum, where all modules of the model were trained simultaneously. Our results indicated that only the \textit{step-by-step} curriculum made the model learn multi-digit addition successfully as reflected by replicating empirical effects of carry-over and problem size. These findings highlight the importance of structured, incremental learning in both cognitive modeling and education.

The transition from descriptive regularities to prescriptive expectations is linked to the inherence heuristic (a cognitive shortcut attributing observed associations to inherent properties), reinforcing the perception of internal characteristics as defining features of social categories and contributing to gender stereotype endorsement. We investigated how inherent reasoning and moderating factors (i.e. generics and individual characteristics) influence the endorsement of gendered activities. Using a 3 (framing: generics vs. "most" vs. "some") × 2 (typicality: typical vs. countertypical gender associations) design, 241 French participants provided descriptive and prescriptive judgments about gendered associations, with justifications coded for inherence. Results showed that generic statements increased prescriptive judgments and reliance on inherent reasoning compared to "most" statements. Inherent justifications increased prescriptive judgments for typical and reduced them for countertypical gender associations. Inherent reasoning fully mediated the effect of generics on prescriptive judgments. These findings underscore the role of language and cognition in sustaining normative gender expectations.

In cognitive neuroscience, there has been growing interest in adopting sequential sampling models (SSM) as the generative choice function for reinforcement learning (RLSSM) to jointly account for decision dynamics within and across trials. However, such approaches have been limited by computational tractability due to lack of closed-form likelihoods for the decision process or expensive trial-by-trial evaluation of complex reinforcement learning (RL) processes. We enable hierarchical Bayesian estimation for a broad class of RLSSM models, using Likelihood Approximation Networks (LANs) in conjunction with differentiable RL likelihoods to leverage fast gradient-based inference methods including Hamiltonian Monte Carlo or Variational Inference (VI). To showcase the scalability and faster convergence with our approach, we consider the Reinforcement Learning - Working Memory (RLWM) task and model with multiple interacting generative learning processes. We show that our method enables accurate recovery of the posterior parameter distributions in arbitrarily complex RLSSM paradigms, and moreover, that in comparison, fitting data with the equivalent choice-only model yields a biased estimator of the true generative process. Moreover, leveraging the SSM with efficient inference allows us to uncover a heretofore undescribed cognitive process within the RLWM task, whereby participants proactively adjust the decision threshold as a function of WM load.

The apparent redundancy of grammatical gender systems presents a puzzle to which information theory offers a solution: since nouns are the least predictable part of language, dividing nouns into semi-arbitrary classes reduces the uncertainty associated with them. Corpus studies show both how gendered articles make nouns more predictable, and how in languages that lack noun class, prenominal adjectives serve a similar function. This raises the question of whether language users are sensitive to this information and actually employ it in communication. In an elicitation study, we manipulated the contextual information provided by gendered articles to German speakers, and compared their behavior to English speakers (whose articles are always uninformative). When German articles were uninformative, German and English speakers produced prenominal adjectives at the same rates. However, when articles were informative, German prenominal adjective production decreased. These results illustrate how languages use both articles and prenominal adjectives to support communicative efficiency.

Manga has gained global popularity, yet how its visual elements, such as characters, text, and panel layouts, reflect the uniqueness of individual works remains underexplored. This study investigates the contribution of panel layouts to manga identity through both quantitative and qualitative analysis. We trained a deep learning model to classify manga titles based solely on facing page images, and performed ablation experiments by removing characters and text, retaining only panel frame structures. Using 10,122 images from 104 works in 12 genres in the Manga109 dataset, we demonstrate that panel layouts alone enable high-accuracy classification. Grad-CAM visualizations further reveal that the models focus on layout features such as size, spacing, and alignment. These findings suggest that panel layouts encode work-specific stylistic patterns and support visual narrative comprehension, highlighting their role as a key component of manga's visual identity.

Large Language Models (LLMs) have been extensively employed in Legal Judgment Prediction (LJP) in recent years. However, existing LLM-based methods often fail to effectively simulate the cognitive processes of human judges, particularly in keyword extraction, leading to suboptimal predictions. Inspired by cognitive science, we propose KnowJudge, a knowledge-driven framework, which explicitly models the cognitive process of legal decision-making, leveraging keyword extraction and precedent-based enhancement to guide LLMs in structured legal reasoning. By integrating external legal knowledge tailored to fact descriptions, it refines keyword identification and selects relevant case precedents, thereby mitigating ambiguity in legal judgment. Unlike conventional methods that rely on fine-tuning, KnowJudge improves performance purely through cognitive-process simulation. Experiments on five benchmarks show that KnowJudge outperforms baseline methods, including both general and legal LLMs.

We report a study examining developmental changes in perceptions of disagreements among speakers who use generics to describe contextually-restricted or unrestricted regularities. Sixty-five adults and 222 5-12-year-olds reacted to generic claims from speakers who attributed ostensibly contradictory attributes to a biological kind ("Xs are striped"; "Xs are spotted"). Crucially, we manipulated the scope of each speaker's claim, described as restricted to a specific context, or unrestricted. Participants assessed faultless disagreement (whether the speakers could "both be right"). Adults were sensitive to contextual restriction: they allowed for faultless disagreement when contextual restrictions mis-aligned, and denied it when both speakers restricted to one context. Young children demonstrated striking partial competence in faultless disagreement judgments much earlier than prior developmental literature suggested. This is the first study to document faultless disagreement between differentially-restricted generics, both in adults and in children. We discuss the developmental trajectory, and implications for social functioning and learning.

Current measures of cognitive conflict in experimental settings focus either on whole trial-level measures, such as reaction time and proportion of cohort disagreement, or intrusive on-task measures such as think-aloud paradigms. Consequently, granular within-trial measurements of the experience of cognitive conflict have been missing from the literature, and consequently, from formal models and theories of decision making. By combining the recently proposed switch paradigm for measuring cognitive conflict with on-task eye-tracking, we ask one such theoretical question: is the experience of cognitive conflict different when choices have clear normative answers and when they don't? Our results answer this question affirmatively and characterize it quantitatively by means of gaze signatures for both classes of experience of cognitive conflict.

The resilience of social networks hinges on identifying users whose departure causes cascading collapse, influenced by both topology and social cognition, such as spatial relationship constraints. Existing studies often overlook how cognitive and behavioral factors shape network fragility. This paper introduces a cognitive-computational framework to detect critical collapsed nodes under spatial constraints, using the (k, σ)-core model to integrate social cohesion (k-core) and spatial thresholds (σ). We propose a pruning algorithm leveraging spatial locality for efficient querying of collapsed nodes and formalize the problem of finding optimal collapsed nodes as an NP-hard task. Our greedy heuristic prioritizes nodes with the most significant cascading impact, similar to human strategies in crises. Experiments on eight real-world networks show our model outperforms topology-only baselines in predicting collapse patterns, especially in spatially-embedded communities. Our findings highlight how spatial constraints and social cohesion amplify systemic fragility, providing insights for designing cognitively-aligned interventions to boost network resilience, bridging computational analysis with cognitive science.

Whether language is essential, sufficient, or a tool for numerical cognition has been hotly debated. Here, we investigate the influence of language on quantity representations by comparing embeddings from vision-only Transformer models (ViTs) and vision-language models (VLMs) exposed to image pairs depicting either the same or different stimulus quantities. If linguistic exposure stabilises quantity representations, VLMs should produce more distinct representations for image pairs with differing numerosity and more similar representations for those with identical numerosity than ViTs. We operationalized this as the variance in Cosine Similarity in response to either categorical (same/different) or continuous differences in stimulus numerosity. We find that VLMs and ViTs are sensitive to the numerosity of visual stimuli, that this sensitivity increases with layer depth, and that VLMs exhibit slightly more sensitivity to image numerosity than ViTs. This work provides initial support for the claim that linguistic exposure can, in principle, stabilise quantity representations.

The diversification of attacks jeopardizes cyberspace's normal operation. This paper proposes a new Harris Hawks Optimization Based on Multiple Strategies (MS-NHHO), inspired by humans' limited cognitive load, collective decision-making, and dynamic learning mechanisms for processing complex information. This paper utilizes the elite chaos reverse learning strategy to improve the algorithm's convergence speed and population diversity. Then, the dynamic adaptive weights are introduced into the escape energy decline mechanism to improve the algorithm's global exploration and local exploitation ability. Finally, the Gaussian random walk strategy enhances the algorithm's anti-stagnation ability. The experimental results confirm the usefulness of the three optimization strategies. Meanwhile, MS-NHHO exhibits satisfactory performance in terms of computational cost, detection performance, and efficiency in several scenarios.

How to improve the persuasiveness of online requests is crucial to achieve acceptance and foster positive social relationships. The effectiveness of persuasion strategies and the influence of the sequence in which these strategies are applied have been demonstrated in the literature. However, existing research has largely overlooked the importance of linking the sequential persuasion strategies to the specific purpose. In this study, we first employ a few-shot Iterative Collaboration Method (ICM) to identify the purpose of the online requests, referencing human needs as well as the persuasion strategies used. Then, the sequential patterns of persuasion strategies supporting respective purposes are mined. Finally, Large Language Models (LLMs), incorporating the identified effective sequential strategies are used to rewrite the original requests. The results indicate that the sequence of strategies used for different purposes, can significantly increase the level of persuasion. The code and dataset can be found at https://github.com/phillip2f/Seq-Strategies-and-Purpose.

Aging is associated with task-specific changes in eye movements, and thus eye movements during mind wandering (MW) in older adults may differ from young adults. We showed that changes in number of fixations or fixation duration were associated with MW in young adults when searching for information but not older adults, possibly due to aging-related changes in these measures. Similarly, larger variance in pupil diameter change was associated with MW when imagining a scenario in young but not older adults, possibly related to aging-related affect stability. In contrast, lower eye movement consistency was associated with MW when implementing well-learned visual routines in older but not young adults, possibly related to their higher susceptibility to MW interferences. Reduced joint attention with another participant was associated with MW for tasks involving clearly defined strategies for both young and older adults. These results have important implications for monitoring task engagement through eye tracking.

The present study examined how participants (N = 298) assessed causality, blameworthiness, foreseeability, and counterfactuality of an AI or human therapist, across three levels of empathy, in comparison to their supervisor and a recommending clinician. We found that participants judged the human therapist as more causal and blameworthy than their supervisor when medium or low empathy levels were displayed, whereas no difference emerged between the judgments of the AI therapist and its supervisor across all of the empathy levels. Additionally, participants did not differentiate causality and blameworthiness between the AI and human therapists, regardless of the empathy level. However, they did perceive the human therapist as foreseeing the outcome more than the AI therapist in the medium and low empathy levels. Qualitative analysis revealed that participants considered the directness of the causes to the outcome, counterfactual reasoning, and inherent limitations of AI when making judgments.

Many languages mark either accusative case (for objects of transitives) or ergative case (for subjects of transitives), but some `split ergative' languages mix the two systems depending on the type of nominal. It has been noted that these languages tend towards marking the less frequent case for each nominal type. This raises the question of what mechanism could underlie the emergence of such an efficient system. We propose a model that can provide an explanation, based on a simple reinforcement learning framework and simple assumptions about asymmetries between the kinds of nominals (e.g., pronouns vs. full noun phrases) that appear in subject vs. object position.

Most of the studies so far overlooked the role of inner speech in action, especially new actions. We conducted a behavioral experiment asking participants to observe videos to acquire two actions. In the experimental group, participants performed an articulatory suppression task, continuously repeating a syllable, so as to interfere with the inner speech. In the control group, participants were requested to continuously tap their middle finger on the table. We hypothesized that, interfering with inner speech, participants could not provide themselves with instructions about how to perform the actions, consequently impacting their ability to acquire them. The results confirmed the hypotheses. Compared to the Dual-task Control Group, the Articulatory Suppression Group is overall more impaired in action acquisition and, in part, in motor performance quality. These results confirm the role of inner speech in cognition, providing new evidence about its function in action learning and execution.

Listeners can rapidly adapt to non-native accented speech, yet the underlying mechanisms remain debated. This study examines whether accent adaptation reflects adjustments in phonetic representations or shifts in decision-making processes. Using a pretest-exposure-posttest paradigm, we examined native English listeners' perception of the Mandarin-accented /θ/-/s/ contrast across two exposure conditions: exposure to Mandarin-accented sentences (Experiment 1) or to pure tones (Experiment 2). In both experiments, listeners showed increased acceptance of ambiguous /θ/ and /s/ tokens when they formed real words, suggesting that adaptation stems from changes in lexical decision criterion reinforced through task repetition rather than accent exposure alone. Additionally, we observed evidence suggestive of rapid within-test distributional learning from limited trials. Our findings support the notion that listeners lower lexical decision criterion when processing accented speech, while also demonstrating remarkable adaptability to novel accent features even with minimal exposure.

Narrative interaction plays an important role in shaping people's beliefs and behaviors both online and in the offline world. We present an experiment examining whether a simple intervention of effect prompting---asking participants to list the effects of complex events---impacts the narrative framing of their networked interactions. After reading a text-based narrative about the Fukushima nuclear disaster, participants in a fully connected network interacted with their neighbors and received rewards for submitting hashtags that matched those of their network partners. Half of the groups received an \textit{effect-prompting} intervention, which shifted participants toward producing more effect-oriented hashtags during networked interactions. We found that the effect-prompting instruction influenced the hashtags participants generated during the network interaction. However, the extent of this shift in hashtags depended on how likely the group was to achieve global coherence.We also examined these dynamics with networks of interacting large language model (LLM) agents using Llama-3.1-8B-Instruct. The study highlights how language-based prompting can subtly shift the narrative framing of online communication.

Converging evidence suggests that languages are shaped by a drive for efficient communication. In particular, it has been shown that languages efficiently compress meanings into words via the Information Bottleneck (IB) principle in domains ranging from visual percepts, such as colors and objects, to non-visual high-level concepts, such as pronouns and number. These domains, however, capture only static elements described by adjectives, nouns, function words, or grammatical markers, leaving open the question of whether the same theory could also apply to verb meanings, which often refer to dynamical aspects of the environment. We address this question by considering locomotion verbs (e.g., walk, run, and jump) across four languages (English, Dutch, Spanish, and Japanese). We show that locomotion verb meanings across languages are shaped by pressure for efficiency, which resonates with similar findings in other domains and suggests that the IB principle may apply more broadly across the lexicon. Our results also open a new avenue for future work to explore whether semantic categories of actions are rooted in a strictly perceptual representation, or perhaps in motor and functional representations as well.

Understanding human perception of visual complexity is crucial in visual cognition. Recently (Shen, et al. 2024) proposed an interpretable segmentation-based model that accurately predicted complexity across various datasets, supporting the idea that complexity can be explained simply. In this work, we investigate the failure of their model to capture structural, color and surprisal contributions to complexity. To this end, we propose Multi-Scale Sobel Gradient which measures spatial intensity variations, Multi-Scale Unique Color which quantifies colorfulness across multiple scales, and surprise scores generated using a Large Language Model. We test our features on existing benchmarks and a novel dataset containing surprising images from Visual Genome. Our experiments demonstrate that modeling complexity accurately is not as simple as previously thought, requiring additional perceptual and semantic factors to address dataset biases. Thus our results offer deeper insights into how humans assess visual complexity.

Convolutional neural networks (CNNs) not only recognize objects with high accuracy, but also acquire from images abstract statistical concepts such as numerosity and correlations. However, it remains unclear whether the CNN architectures implement inductive biases that mimic human biases in statistical judgments. In this paper, we examined whether VGG-19 models, a popular CNN architecture, that are trained to make correlation judgments from scatterplots display human-like biases. In comparisons between model predictions and human data, we found that there was a high correspondence between human biases and machine biases in VGG-19 models. Using explainable AI visualization with saliency maps to unpack the regions on which VGG-19 rely to make correlation judgments, we found that the late layers of the model tend to focus on regions similar to human participants' fixation distributions as captured by eye tracking. We further demonstrate that such models were nearly sufficient to predict human data at an accuracy level rivaling the state-of-the-art model trained on human data in three large-scale correlation discrimination datasets. Our results suggest that VGG-19 models may employ strategies that are similar to those used by human participants for statistical judgments from visual graphs and, therefore, pave the way to address human cognitive biases in visualization-based statistical judgments through the lens of deep neural networks.

One explanation for how people can plan efficiently despite limited cognitive resources is that we possess a set of adaptive planning strategies and know when and how to use them. But how are these strategies acquired? While previous research has studied how individuals learn to choose among existing strategies, little is known about the process of forming new planning strategies. In this work, we propose that new planning strategies are discovered through metacognitive reinforcement learning. To test this, we designed an experiment to investigate the discovery of new planning strategies. We then presented metacognitive reinforcement learning models and demonstrated their capability for strategy discovery as well as show that they provided a better explanation of human strategy discovery than alternative learning mechanisms. However, when fitted to human data, these models exhibited a slower discovery rate than humans, leaving room for improvement.

This paper presents a framework for understanding metacognition as a distinct domain of skill, drawing on established research in motor and cognitive domains. It proposes that metacognitive expertise shares key characteristics with other skill domains, including goal-directed action, hierarchical organization, declarative and procedural knowledge, and automatization. By integrating theoretical and empirical insights, this paper aims to establish a comprehensive model of metacognitive skill development, with implications for research and practical applications in education, therapy, and beyond.

Visual narrative comprehension is crucial in today's visually driven world, yet research on aging and this ability remains inconclusive. The role of stress in this process is largely unexplored, despite stress becoming an ever-present factor in modern life. This online study investigated how acute stress affects visual narrative comprehension in younger (N = 203, 18-57 years) and older adults (N = 212, 60-85 years). Participants were assessed under stress and neutral conditions. Stress was induced through timed mathematical and logical tasks with social stress elements. Pictorial stories, consisting of three panels with the second intentionally left blank, were presented. On the next page, participants judged whether the given inference for the blank panel was correct. Results showed that acute stress impaired comprehension and confidence in younger adults, while older adults remained unaffected. These findings suggest that with age and experience, individuals develop more differentiated event schemas, enhancing resilience to stress.

Assessing learning motivation and strategy (LMS) in specific situations can more accurately reflect students' self-regulation learning ability. However, traditional assessment methods, such as subjective evaluations and self-reports, are time-consuming, burdensome, and not well-suited to the dynamic nature of situational assessments. To address this, we presented the LLM-based agents, which enable intelligent generation of situational tasks and interactive assessment. Specifically, Master defines the theme and storylines, Designers generate situational tasks, Evaluator reviews the content quality, and Interactor controls the interactive assessment with users. The results of a user study with 97 university students demonstrated the reliability and validity of our approach and the significant enhancement of the user experience. The results further clarify the relationship among indicators of LMS. This study provides a novel paradigm and solution for situational assessment of LMS and offers valuable theoretical insights for intervention research targeting related indicators.

Recently, AI-generated content (AIGC) technologies have made remarkable advancements, even achieving superhuman performance across various domains. However, few previous studies have investigated its impact on emotion-focused therapy with artistic content, e.g., music. In this paper, we conducted an EEG experiment to explore the effects of generative music on emotion-focused music therapy based on the ISO principle. This experiment compared AI-generated and human-created music regarding the changes in participants' valence and arousal following negative emotion induction with the ISO principle adherence and non-adherence. The results show that generative music, with its harmonic consistency and simple rhythm, is more effective in supporting positive emotions and improving temporal lobe activity. Besides, the therapeutic effectiveness of generative music adhering to the ISO principle has also been validated. This study highlights the distinct emotional and neural mechanisms of AI-generated music, offering valuable insights into future AI-powered emotion-focused therapy strategies.

Humans invented reading and have passed down this complex skill across generations through language. This study provides empirical evidence of the neural mechanisms underlying bottom-up (related to high-order linguistic structure) and top-down (related to next-word predictability) processes, which interact to guide comprehension during reading. While previous studies have focused on either the N400 effects of predictability or lexical categories, research on how predictability influences N400 responses across different lexical categories is limited, mainly due to constraints in publicly available datasets. Here, we examine how predictability influences brain responses, recorded at millisecond resolution using electroencephalography (EEG), with a focus on the N400 time window (300-500 ms post-stimulus) across different lexical and grammatical categories. Our results indicate that significant differences in N400 responses between high and low cloze probability levels were more pronounced for content words than function words. Among the two primary content categories, verbs exhibited greater N400 differences than nouns, while nouns carried more distinct information about their predictability than verbs. Moreover, we demonstrate that the decoding technique is more effective than the event-related potential (ERP) traditional analysis in capturing more detailed and distinct representations of cognitive processes over time.

Recommender systems personalize content delivery based on user's interaction history. However, not all clicks result from deliberate decisions—many arise from intuitive reactions. Inspired by the dual process theory, we argue that intuitive clicks are primarily driven by System 1, reacting to superficial cues, while analytical clicks involve deeper processing by System 2, considering the semantic meaning and long-term preference. However, existing models overlook these cognitive mechanisms. To address this, we propose DualRec, a novel recommendation method that models both intuitive and analytical behaviors. DualRec encodes items using language models, leveraging shallow layers for superficial understanding (System 1) and deep layers for semantic comprehension (System 2). It employs Transformer-based encoders with two attention mechanisms to capture intuitive "fast" and analytical "slow" click patterns. A learnable fusion layer balances these behaviors. Extensive experiments demonstrate that DualRec outperforms existing methods and highlights the importance of integrating both cognitive processes in recommendations.

The process by which people adjust, enrich, and revise their understanding of word meanings over time – so-called ‘Slow Mapping' – has often been overlooked, particularly in terms of how a computationally bounded learner might approach such a task. To address this gap, we propose a process model of incremental word-meaning induction. This proposal is inspired by recent work on concept and theory change grounded in a probabilistic language of thought (pLOT). We focus on the problem of fixing the meanings of words from usage examples, taking kinship terms as our test domain. We frame word meaning induction at a computational level as a program induction problem, and hypothesize that individual learners search for possible meanings as evidence arrives via a mutative Markov-Chain Monte-Carlo search scheme. We show this idea provides a better description of how participants' generalizations and tentative definitions of alien kinship words shift as evidence arrives, outperforming normative accounts and other baselines.

This paper probes the influence of a particular kind of domain-general cognitive bias in first language acquisition with the aid of computational models. We introduce a novel task: inducing functional categories from morphemically tokenised sentences, and supply a manually annotated dataset of English child-directed speech (CDS). We operationalise a widely assumed type of cognitive bias, "less-is-more", as three computational principles—ordering input, gradually increasing model complexity, and priming the learner—and develop a theory-neutral experimental setup to evaluate their impact on functional category induction. Our experiments with CDS demonstrate that models incorporating reflexes of "less-is-more" outperform the purely statistical baseline. As part of our exploration of ordering effects, we employ the morpheme acquisition order proposed by Brown (1973) and, for the first time in literature, present statistical evidence that Brown-compliant orders outperform non-Brown-compliant ones.

During human communication, speakers balance informativeness and effort by tailoring their language to their audience. Large Language Models (LLMs) appear human-like in their communication and succeed at some tasks thought to involve social reasoning about interlocutors (in humans). Here, we tested audience design in LLMs using tasks modeled on (Isaacs & Clark, 1987). In Experiment 1, replicating findings with humans, LLMs produced longer responses when producing descriptions of pictures from a city while addressing an audience unfamiliar with that city and used more proper nouns when addressing a familiar audience. In Experiments 2 and 3, similar to previous findings with humans, LLMs used fewer words to describe pictures over the course of a multi-turn interaction. However, this pattern appeared to be sensitive to whether the user prompts also got shorter across turns, suggesting that efficient audience design in LLMs reflects patterns in training data and reinforcement learning, rather than an inherent drive towards least effort.

Memory has traditionally been studied in well-controlled laboratory environments, which, while effective, do not fully capture the range of dynamics and behaviors shown in real-world contexts. To address this gap, we propose using summarization as a novel task to study memory recall in naturalistic settings. We argue that a key component of summarization is the ability to represent and retain information from the original material. Inspired by approaches in the free recall literature to analyze temporal dynamics of memory recall, such as how recall begins and transitions to subsequent items, we analyzed the temporal dynamics of summary patterns. Using three publicly available summarization datasets and a naturalistic narrative recall dataset, we found alignments between the summarization patterns and established free recall patterns, including primacy, recency, temporal contiguity, and the effect of list length. These results support that summarization involves processes of memory recall and open up opportunities to use summarization as a naturalistic task to study memory recall in the future.

PLGAN is a generative model-based framework for personalized knowledge tracing, designed to explore and model individual learning preferences. By integrating a Personalized Attention Mechanism (P-Attn), PLGAN effectively extracts learners' distinct learning patterns and behavioral tendencies, addressing the limitations of traditional knowledge tracing models that assume homogeneous learner groups. Unlike conventional approaches, PLGAN dynamically adjusts the weighting of behavioral features, enabling a more nuanced representation of learning preferences and improving the accuracy of knowledge state predictions across diverse learning environments. Experimental results on multiple public datasets demonstrate that PLGAN achieves an average performance improvement of 3.5% in knowledge tracing tasks. Furthermore, the generative nature of PLGAN enhances its generalization and robustness, effectively capturing individualized learning dynamics. This work advances the study of personalized learning by leveraging generative models to model and analyze learner behavior, providing a cognitively informed approach to knowledge tracing.

Humans naturally interpret numbers non-literally, effortlessly combining context, world knowledge, and speaker intent. We investigate whether large language models (LLMs) interpret numbers similarly, focusing on hyperbole and pragmatic halo effects. Through systematic comparison with human data and computational models of pragmatic reasoning, we find that LLMs diverge from human interpretation in striking ways. By decomposing pragmatic reasoning into testable components, grounded in the Rational Speech Act framework, we pinpoint where LLM processing diverges from human cognition --- not in prior knowledge, but in reasoning with it. This insight leads us to develop a targeted solution --- chain-of-thought prompting inspired by an RSA model makes LLMs' interpretations more human-like. Our work demonstrates how computational cognitive models can both diagnose AI-human differences and guide development of more human-like language understanding capabilities.

Many causal relations can be represented by continuous functions that map inputs to outputs. Can young children learn continuous causal functions and generalize them from observed data to new scenarios? We found that 4- and 5-year-olds can represent continuous functions with different abstract forms. After observing a few input-output pairs, children can accurately infer positive linear and step functions by predicting the outputs of novel input values. They also have emerging knowledge of negative linear and triangular functions. While children do not yet make consistently accurate predictions for these functions, they can distinguish these functions from the positive linear function and show inferential patterns that are consistent with the respective functions. Like adults and older children, preschoolers show an inductive bias towards positive linear functions. Their understanding of negative linear functions--which strongly requires inhibiting this inductive bias--improves with age.

Belief bias in syllogistic reasoning occurs when individuals' agreement with a conclusion influences their assessment of its logical validity. While this effect has been widely studied in domains such as politics and personality-related reasoning, its role in evaluating arguments about sexual violence remains underexplored. In a pre-registered study, we examined whether participants' sexist beliefs and cognitive reflection influenced their ability to assess the validity of syllogisms related to sexual violence. Participants (N = 104) completed a syllogistic reasoning task with sexism-supportive, sexism-challenging, and neutral syllogisms, as well as the Ambivalent Sexism Inventory and a Cognitive Reflection Test. The results indicate that when evaluating such syllogisms, participants' beliefs play a significant role. People tend to perceive syllogisms as logically valid if the conclusions align with their beliefs and as logically invalid if the conclusions contradict their beliefs. Furthermore, cognitive reflection moderated belief bias effects, but only for sexism-supportive syllogisms. These findings highlight the extent to which reasoning about gender and sexual violence is shaped by preexisting beliefs and suggest that cognitive reflection may help mitigate some bias-driven reasoning errors.

How does linguistic telicity map onto mental representations of events? Recent work suggests considerable flexibility in how people mentally represent temporal event structure, yet we know little about how linguistic cues modulate these representations. We investigated how different forms of quantization correspond to event construal using a novel experimental paradigm that bridges event perception and linguistic processing. Participants first learned to distinguish bounded from unbounded events, then categorized sentences varying in quantization strength. Our results revealed a systematic relationship between linguistic form and event representation: strongly quantized expressions ("drink one beer") reliably corresponded to bounded event construal and activity descriptions ("did some writing") to unbounded interpretations, while bare plurals ("drink beers") showed genuine flexibility in interpretation. This graded pattern indicates that temporal boundedness in cognition operates along a continuum, with linguistic cues providing weighted probabilistic constraints on event representations. The findings demonstrate how different linguistic forms correspond to varying degrees of flexibility in event understanding, contributing to our knowledge of how language interfaces with event cognition.

Automatic emotion recognition using electroencephalography (EEG) signals has garnered significant attention in recent years. While multi-source domain adaptation methods provide a promising framework for cross-subject emotion recognition, the distributional discrepancies among different source domains often result in negative transfer. To address these challenges, we propose a two-stage Dynamic Domain Selection and Pseudo-label Refinement (DDSPR) model. In the first stage, we introduce a novel Dynamic Domain Selection (DDS) module and an Agent Domain Adaptation Strategy (ADAS) to dynamically select and align source domains. In the second stage, a confidence-based pseudo-label correction strategy is employed to refine target domain labels and mitigate noise. We evaluate the proposed model through cross-subject experiments on the SEED and SEED-IV datasets, achieving accuracies of 91.50% ± 7.05 and 78.05% ± 13.56, respectively.The results demonstrate its effectiveness in emotion recognition performance.

When are in-group linguistic conventions opaque to non-group members (teen slang like "rizz") or generally interpretable (regionalisms like "roundabout")? The formation of linguistic conventions is often studied in iterated reference games, where over repeated reference to the same targets, a describer--matcher pair establishes partner-specific shorthand names for targets. To what extent does the partner-specificity of these linguistic conventions cause them to be opaque to outsiders? We use computational models and experiments with naïve matchers to assess the opacity of descriptions from iterated reference games. Both human matchers and the computational model perform well above chance, suggesting that conventions are not fully arbitrary or opaque, but reflect aspects of shared semantic associations.

When observing an action, how do people intuitively determine if it has communicative intent, such as in the case of pantomimes? We focus on two alternative theories: one suggests that instrumental intention competes with communicative intention, where the weaker the former is, the stronger the latter is; the other proposes that instrumental intention is nested within communicative intention, where the presence of the former facilitates the latter. To test these theories, we introduced the concept of contextual affordance, which manipulates the degree to which the instrumental components of an action are revealed. Through behavioral experimentation, we found a non-linear relationship between contextual affordance and communicativeness rating: partial affordance–providing implicit cues to an action's instrumental component without being fully rational–elicited the strongest perception of communicative intention, whereas full affordance or no affordance resulted in a weaker perception of communicative intention. Our study provides a novel definition of communicative intention and reveals that recognizing the instrumental goal and perceiving the suboptimality in achieving it work together to create a strong communicative signal. This work represents a step toward developing an integrated theory of pantomimes, specifying how the rationality principle can be applied to serve multiple purposes simultaneously.

Researchers have started seriously considering the epistemic issue of whether and when we can claim an artificial intelli- gence (AI) has developed machine consciousness. Most cog- nitive theories of consciousness employ a functional character- ization of the property of consciousness. That is, they are com- mitted to an account of consciousness as a rule-governed pro- cess over mental states. Some cognitive scientists concerned with AI advocate an epistemically behaviorist approach to ma- chine consciousness; however, such approaches taken ontolog- ically, systematically fail to satisfy reasonable intuitions about in what consciousness ought to consist, and taken epistemi- cally, fail to provide sufficient evidence to individuate any in- ternal property, including consciousness, in non-human sub- jects. Therefore, in order to assess consciousness in ways that adequately account for reasonable intuitions as to its proper definition, such that we can reasonably assert the presence of machine consciousness in some AI, it is necessary to propose, test, and revise, functional theories of consciousness.

Prior research has shown that people vary considerably in how they interpret political concepts, a variability often attributed to liberal–conservative differences underlying political polarization. In this study, rather than focusing on the liberal–conservative dichotomy, we considered personality and morality variables as possible predictors of cognitive diversity in subjects' interpretation of political concepts. Participants completed brief personality (HEXACO) and morality (MAC) assessments, followed by a series of association ratings for the concepts of freedom, justice, and authority. We found that certain personality traits and moral dimensions correlate with higher associations between probe concepts. Furthermore, clustering of political inclination on morality dimensions and concept ratings suggested that the latter made a limited contribution to political diversity, only raising the number of clusters from 2 to 3. Keywords: personality; morality; conceptual diversity; freedom; justice; authority

The single-celled protist Stentor roselii has long been observed to exhibit complex decision-making behaviors, yet existing machine learning and classical computational models have struggled to replicate its actions. In this paper, we propose a novel quantum-statistical framework to model S. roselii's behavioral responses to environmental stimuli. By leveraging quantum circuits with amplitude dampening and memory effects, we construct a quantum behavioral model that captures the probabilistic and hierarchical nature of S. roselii's decision making. Our results suggest that quantum statistical theory provides a powerful tool for representing and simulating biological decision processes.

The role of self-generated movement in memory retrieval has been demonstrated in enactment paradigms. However, in the context of object-location memory, the impact of action during learning has not yet been investigated, despite the ecological relevance of such behaviors. In the current project, we present new evidence that actively placing an object in a target location during learning leads to more precise, and faster, subsequent recall of the object-location associations than simply observing this placement. We further demonstrate differences in object- location memory depending on the category of stimuli that participants are engaging with by showing that images of objects with high manipulability are placed more precisely, more quickly, and more directly (mouse-tracking) than images of objects with low manipulability. We suggest that these latter differences are due to the motor information implicitly activated during processing of high manipulability items, and reflect the embodied nature of concepts. Although both enactment and manipulability impacted object-location recall, they did not interact. This research extends findings on enactment to associative encoding processes, and informs our understanding of the relationship between enactment and embodiment in human memory.

Instruction tuning (IT) has been a fruitful technique for aligning Large Language Models with human preferences. However, the linguistic implications of IT remain unclear. In two experiments on coreference and coherence biases in the context of Implicit Causality, we investigate how IT modulates these discourse biases in relation to model size. Our results show that IT interacts with model size -- instruction-tuned models display enhanced coherence biases and more human-like coreference patterns, sometimes exceeding human performance. However, this effect appears size-dependent, suggesting that IT causes some linguistic patterns to emerge that are dormant in the respective foundation models.

Assessing attention is essential for optimizing performance. This study identifies a single-channel EEG biomarker based on cognitive load to estimate Task Resolution Time (TRT). Thirty-seven chess players were recorded with an 8-channel EEG headset while solving chess problems under two conditions: distracting noise (65 dB) and ambient noise (40 dB). Participants were grouped by chess expertise (ELO rating), and cognitive load was measured via theta (4-8 Hz) power on the C4 channel. EEG signals underwent preprocessing with a bandpass filter, Artifact Subspace Reconstruction (ASR), and Independent Component Analysis (ICA). Power estimation (Welch) was normalized to a resting 30-second Eyes Open (EO) period. TRT analysis indicated shorter engagement times and slightly lower performance in novices under noise, while experts remained relatively stable, possibly due to better cognitive resilience. This biomarker could be further integrated into portable EEG systems for real-time neurofeedback in educational and workplace settings.

Children's understanding of privacy develops as they navigate both physical and digital spaces. This study examines how children aged 3- to 13- years-old conceptualize privacy through their drawings, analyzing data from the Privacy Illustrated dataset. We explored two key dimensions: regulation (mechanisms controlling access, such as doors) and content (what children consider private, such as bedrooms or intellectual property). Our findings suggest that as children get older, they are more likely to view privacy as something that can be actively managed using physical barriers and control mechanisms. In contrast, younger children often depicted privacy as simply being alone. Content-related depictions remained relatively stable across ages, though older children included more abstract ideas, such as digital privacy. This study provides a novel framework for examining privacy development, highlighting distinct but interrelated dimensions of privacy.

Human learning relies on specialization---distinct cognitive mechanisms working together to enable rapid learning. In contrast, most modern neural networks rely on a single mechanism: gradient descent over an objective function. This raises the question: might human learners' relatively rapid learning from just tens of examples instead of tens of thousands in data-driven deep learning arise from our ability to use multiple specialized mechanisms of learning in combination? We investigate this question through an ablation analysis of inductive human learning simulations in online tutoring environments. Comparing reinforcement learning to a more data-efficient 3-mechanism symbolic rule induction approach, we find that decomposing learning into multiple distinct mechanisms significantly improves data efficiency, bringing it in line with human learning. Furthermore, we show that this decomposition has a greater impact on efficiency than the distinction between symbolic and subsymbolic learning alone. Efforts to align data-driven machine learning with human learning often overlook the stark difference in learning efficiency. Our findings suggest that integrating multiple specialized learning mechanisms may be key to bridging this gap.

Large-scale annotated data is essential for age prediction in social media, yet obtaining such data is costly. Age is a key psychological and cognitive marker influencing communication and social behavior. Understanding age-related patterns in online interactions can provide insights into cognitive development and identity formation. To address data limitations, we propose a semi-supervised multimodal regression model leveraging Transformer-based variational autoencoders to infer age from textual and social features. This approach aligns with cognitive and social science theories on age-related behavioral patterns. Our framework effectively utilizes unlabeled data, reducing annotation dependency while enhancing predictive accuracy. Empirical results demonstrate superior performance over traditional classification and supervised baselines, advancing interdisciplinary research in age inference and online behavior modeling.

Our knowledge of a topic such as mathematics is reliant upon the hierarchies of chunks we build in our memories. The time course of knowledge-based tasks, such as the transcription of algebraic formulas, provides rich signals that reflect the structure of the chunk hierarchies being processed. By building Gaussian Mixture Models, this paper provides evidence that decomposing the overall dis-tribution of pauses between actions in a sequential task can give meaningful characterizations of the structure of the chunk hierarchy. We also examine whether individual competence in mathematics can be measured using a metric derived from the models.

With the rise of Large Language Models (LLMs), interest in simulating interaction dynamics has grown, raising questions about their validity as cognitive models of human discourse. While extensive research focuses on their performance in various applications, we aim to quantify LLM conversational processes akin to traditional human studies. By analyzing how convergence entropy evolves across different conversational tasks, we propose a framework for quantitatively assessing LLMs' ability to exhibit specific features. This approach offers a pathway to characterizing LLMs for agent-based modeling and broader discourse analysis.

This study examines how prior information about a partner affects the sense of joint agency during a collaborative task. Previous research suggests that perceiving a partner as human enhances this sense. However, the effects of prior instructions remain unclear. We designed a 2 × 2 factorial experiment in which participants were told their partner was human or a program, while their actual partner was either human or a program. In the experiment, each participant and their partner jointly controlled a cursor to trace a circle. Additionally, we measured the sense of agency to compare it with joint agency. The data revealed that the instructional factor only influenced the sense of joint agency, while the actual partner factor solely affected the sense of agency. These findings suggest that beliefs about the partner prior to interaction influence the sense of joint agency, offering insights into the cognitive processes underlying collaborative actions.

Each person develops a unique emotional landscape shaped by their experiences and linguistic-cultural contexts, partly personal and partly shared with others. This enables personally unique emotional experiences while maintaining shared understanding. This work aims to advance a framework for investigating what's shared and distinct across individuals, beginning with linguistic communities as an essential level of analysis, using English and Korean speakers as our case study. We examined how emotion concept representations differ between English and Korean speakers using representational similarity analysis and network analysis. English and Korean speakers' judgments of pairwise similarity between 57 emotion concepts evidenced both substantial shared structure and language-specific patterns (Spearman's � = 0.72, indicating 48% unshared variance). While valence emerged as a key organizing dimension in both languages, network analyses with strength centrality showed distinct patterns for each language. First, the Korean emotion concept network demonstrated higher strength centrality across all emotion concepts than the English network, indicating higher interconnectedness between concepts. Second, high-centrality emotions were predominantly negative in both languages but formed language-specific local networks with different sets of neighboring concepts. The statistics of language usage encode a substantial part of the conceptual structure of emotion, enabling large language models to capture aspects of human emotion. Despite their advanced multilingual capabilities, GPT4-o and Claude-3.5 showed stronger alignment with English speakers' representations, regardless of prompt language. These findings demonstrate that while languages reflect common principles in emotion representation, they shape distinct patterns, with implications for cross-linguistic/cultural emotion understanding and AI system development.

An influential mathematical model of memory, the temporal context model (TCM), posits that we encode items and their associations with temporal context (Howard & Kahana, 2002). Temporal context is conceived of as a recency-weighted av- erage of past experiences. Critically, the model assumes that when an item is retrieved later, the associated temporal con- text is also obligatorily retrieved. Existing evidence for the idea of retrieved temporal context primarily comes from free- recall studies. However, free recall introduces some critical confounds that are difficult to resolve (Folkerts et al., 2018) and also encourages memory strategies that may mimic tem- poral context effects (Hintzman, 2011). To address these con- founds, we investigate temporal context using an image recog- nition task. Schwartz et al. (2005) examined temporal con- text in an image recognition task using a short-list-based ex- perimental design, and found that temporal context influenced recognition performance. Building on this, we use the Natural Scenes Dataset (NSD) to show that reinstating temporal con- text enhances recognition accuracy even across substantially longer timescales. We demonstrate that images that were tem- porally closer during encoding facilitated the recognition of each other. Critically, we show that this influence falls off with temporal distance at encoding only when the temporal context is successfully retrieved, as predicted by TCM. Furthermore, the slope of this temporal gradient increases as a function of the strength of the influence of the retrieved temporal context. These findings extend our understanding of temporal context effects in episodic memory by showing that temporal context is retrieved even in tasks that do not encourage linking between items as a memory strategy.

Self-regulated learning involves making decisions about how to study, but students often choose suboptimal strategies. Our experiment investigated how preferences for instructional methods—video, practice, or both combined—affect learning outcomes. We randomly assigned 130 participants to either receive their preferred method of instruction (honoring initial choice) or a different method (dishonoring initial choice). Contrary to previous research showing preferences for lectures, our participants initially selected practice-based approaches. However, when asked again after instruction, the majority of participants chose the combined approach. This shift in preferences suggests that students may overvalue comprehensive approaches, even when practice alone was equally effective and reduced instructional time by 66%. Whether preferences were honored or dishonored did not significantly affect performance or efficiency, thus control over instructional methods may be less important than the methods themselves. Based on our findings, future research should focus on guiding students to utilize practice to optimize learning efficiency.

Generative artificial intelligence (GenAI) is rapidly advancing, providing innovative tools and methods for a wide range of applications. Among these, Generative Agents, a key domain in GenAI, are valued for their fine-grained definitions and simulations of human-like behaviors. These agents provide new avenues for studying and modeling various domains, including social interactions, education, and cognitive sciences. However, existing works suffer from cognitive dynamics disconnect and affective absence, which hinder researchers from exploring human-related cognitive processes in depth. To address these limitations, we propose EvoAgents, a novel cognitive-driven framework that pioneers the exploration of dynamic personality evolution in Generative Agents. By defining the emotional content of agents and integrating a cyclical personality evolution cycle, EvoAgents represent a significant step toward creating more adaptive and authentic agent behaviors. Comprehensive simulations and evaluations show that EvoAgents achieve superior performance in key automated metrics compared to prior work, while uniquely enabling reasonable and robust personality evolution processes that align with cognitive and psychological expectations. By constructing a new simulation environment, SmallClassroom, based on the EvoAgents framework, we validate the framework's ability to provide deeper cognitive insights into social dynamics, aligning closely with established psychological theories.

There are two competing predictions for how experts will teach novices. According to one account, general rules and principles are better for equipping learners to perform the complex tasks that are characteristic of many domains of expertise. On the other hand, synthesizing a rule can be challenging for an expert, particularly in real-life environments with complex tasks, and learners may not be able to apply the rule to novel cases. In the present study, we examine a sample of six expert-novice pairs in the context of a robot-assisted surgical procedure in a teaching hospital. The nuances of this domain allow us to examine how an expert's decision to teach inductively or deductively varies systematically across task complexity. We find initial evidence that experts in this domain primarily teach using specific case-by-case examples, as opposed to general rules and principles, regardless of the complexity of the surgical task.

Sound symbolism is a linguistic feature that may suggest a non-arbitrary link between sound and semantic content that has been shown to play a role in language acquisition, and potentially language evolution. One sound-symbolic structure known as phonesthemes has been identified in many languages. Phonesthemes are sub-morphemic sound patterns associated with specific meanings more frequently than chance. Cognitive approaches to phonesthemes typically rely on a known example. Phonesthemes are not well attested in Japanese, so this approach is not an easy option. This research provides a route for research into the cognitive effects -- in particular language acquisition -- of Japanese sound symbolism. I present a model for identifying phonestheme-like features in Japanese adapted from a model originally used with English and identify two candidates in Japanese. I also outline methods for empirically testing the psychological reality of these clusters based on existing English and Japanese sound symbolism literature.

Mental health disorders impact nearly one billion people worldwide, yet stigma and insufficient awareness often prevent individuals from seeking timely professional help. The proliferation of social media platforms has introduced new opportunities for detection of mental health conditions, enabling the analysis of user-generated content to identify whether one has a mental disorder. Traditional approaches to this task have largely relied on content-based models, such as n-grams or language embeddings, which are prone to domain-specific biases and often fail to account for the emotional dynamics inherent in mental health expressions. In this work, we propose a novel framework for detecting mental disorders through the analysis of Reddit conversations, integrating both temporal textual data and emotional cues. Our model addresses the limitations of prior methods by explicitly capturing the evolving relationship between textual content and emotional expression over time. Experimental results demonstrate a significant improvement in detection accuracy compared to existing approaches, while ablation studies highlight the critical role of temporal emotional information in enhancing performance. These findings suggest that a more nuanced, emotion-aware approach offers substantial promise for advancing computational mental health diagnostics.

Humans who use a left-to-right writing system often associate smaller numbers with the left side of space and larger numbers with the right. Whether this left-to-right number-space mapping is innate or culturally learned is unclear. Here, we test whether monkeys who lack human cultural practices show a left-to-right number-space mapping. Previous work in monkeys has found mixed evidence on whether monkeys show a left-to-right bias in their number-space mappings. Replicating the methods of Drucker and Brannon (2014), monkeys were trained to touch the fourth circle from the bottom in a vertical array of five circles. Then, they were tested with a horizontal array of five circles. Overall, monkeys showed no preference for the fourth circle from the left compared to the fourth from the right. This suggests monkeys may not have a directionality bias for number-space mappings. Therefore, the left-to-right bias in humans may be due to specific cultural practices.

Financial decisions such as those involved in stock trading should, at least partly, be based on similar features as detecting trends in time-series data. When presented as a purely perceptual task, people's accuracy in detecting trends is generally considered good, but real-world individual investors underperform the market globally. In a series of controlled experiments, we contrast financial decisions to perceptual ones, presenting participants with real-time evolving time series whilst manipulating the reward structure and the context. Our results show that participants' decisions were not affected by trend direction in a classic perceptual decision-making scenario, whilst in a classic trading scenario they performed worse in both speed and accuracy during downward (i.e., bear markets) compared to upward trends (i.e., bull markets). In a final experiment where we carefully controlled the reward structure of both scenarios and the only relevant differentiating factor were the labels of the decisions, we did not find evidence for this difference between scenarios, but participants were slower in the trading scenario. Employing the Drift-Diffusion Model, we found evidence of lower efficiency in the classic trading, compared to the classic perceptual decision-making scenario. Our results provide much-needed insight into the cognitive basis of trading decisions and the general underperformance of real-world individual investors.

Recent advances in large language models (LLMs) have renewed interest in understanding the cognitive mechanisms underlying human language use. In this study, we focus on the Japanese word game ``Shiritori'' as a simple task related to language use, and aim to clarify the cognitive factors involved in its characteristics. To this aim, we model the execution of Shiritori based on the basic memory mechanisms of a cognitive architecture and investigate the characteristics that appear in Shiritori word sequences through simulations with different parameter settings. The results show that word sequences with different properties were obtained depending on the levels of lexical activation, inhibition, and semantic association. To complement the simulation findings, we also conducted a preliminary human evaluation in which participants rated the robot-generated word sequences. By constructing a model that can flexibly control Shiritori behavior, we explore its potential applications as a stimulus for research on human–robot interaction and language acquisition support.

Accumulating knowledge and experience across the lifespan are bound to have an impact on the meaning of words. Here, we investigated this idea using primarily emotional valence of words as a test-case. We used French databases that gather psycholinguistic variables including emotional valence of words, in four groups of individuals including young (18-25; 26-39 years), middle-aged (40-59 years) and older (>60 years) adults. Following the hypothesis that words may display age-related differences in their psycholinguistic properties, we computed linear regressions over all individual words as a function of age-groups. Results revealed notably that between 5 and 10% of words show significant linear changes of emotional valence as a function of age. This pattern highlights the situated and flexible nature of word meanings and suggests that self-relevance of experience affects semantic memory.

Spaced retrieval practice has been repeatedly demonstrated to improve learning, but its implementation is often constrained by real-world time limitations. This study investigated whether, under fixed study durations, learners should prioritize spacing or repetition. Across two experiments (total N = 1589), participants practiced Indonesian vocabulary under four conditions that varied in spacing and repetition. Item difficulty was also manipulated. Results showed that increasing repetitions at the cost of spacing enhanced immediate test performance, particularly for harder items. These findings suggest that spaced retrieval practice is effective only when learners have sufficient prior repetitions to retrieve information successfully. This study highlights the trade-offs between spacing and repetition under time constraints and offers practice insights for optimizing learning strategies.

Object knowledge comprises a virtually limitless feature space. When cued to generate attributes for an exemplar (dog), any feature is possible (has molecules). Yet, some features are definitively favored (tail). We hypothesize a sensorimotor bias in feature generation wherein perceptually salient features upon first pass are more cognitively accessible than abstract, verbally mediated knowledge. We examined the role of sensorimotor dominance in semantic feature generation by yoking concreteness values to cues (N=4436) and features (N=69,284) within the Buchanan et al. (2019) norms. We predict that cues regardless of their concreteness evoke relatively more concrete features (e.g., dogs evoke tails, justice evokes lawyers). The data moderately supported this hypothesis. Feature concreteness increased linearly with cue concreteness (R=.83) but the y-intercept (2.78) indicates that overall, features were more concrete than their cues. We discuss alternate factors (e.g., frequency, familiarity) that may moderate the likelihood that people retrieve tail when cued with dog.

Research on growth mindsets, emphasizing the malleability of intelligence through effort, often highlights their benefits of boosting performance and reducing achievement gaps. Across four studies (N = 785), we investigated the unintended consequences of the growth mindset, hypothesizing that its emphasis on intelligence as controllable would lead to greater blame toward others for intellectual failure, compared to the fixed mindset, which views intelligence as largely innate. Study 1 found that participants primed with the growth mindset assigned more blame for low-difficulty intellectual failures than those primed with the fixed mindset. Study 2 showed that this effect diminished when intellectual failures involved highly challenging tasks. Study 3 highlighted the harm caused by an individual's intellectual failures and found that participants in the growth mindset condition still assigned greater blame than those in the fixed mindset condition. Study 4 explored a possible mechanism, finding that a growth mindset, compared to a fixed mindset, increased blame by leading participants to perceive less effort from the protagonist in the vignettes, even when both conditions were faced with identical intellectual failures. These findings underscore the need for nuanced implementations of the growth mindset.

How do children learn the appropriate scope of linguistic generalizations? One proposal is that prediction error and cue competition enable them to implicitly reduce their uncertainty about the various cues to linguistic patterns. Previous work has employed artificial language studies to test the predictions of error-driven models against the performance of (adult) human participants (Ramscar et al., 2010). A critical prediction of these models - that linear relations between linguistic and environmental cues can critically affect generalization - has received much empirical support. For example, Vujovic et al. (2021) found that suffixing languages supported the learning of discriminating cues, and overgeneralization avoidance, better than equivalent prefixing languages. The current study addresses a limitation of previous studies: the use of unnatural flat distributions, which contrast to the skewed distributions ubiquitous in natural language. Although some of our results are consistent with model predictions, there were divergences. Possible reasons for these are discussed.

Despite growing concerns about the effect of social media en- gagement on people's beliefs and behavior, estimating the ac- tual impact is difficult. Here we present preliminary results from our own isolated social media platform named Magpie Social. In it, participants could interact with each other like typical social media, but we had control over the platform and measured people's beliefs and behavior before and after us- ing it. This allowed us to more closely approximate the eco- logical validity of naturally occurring social-media data, while retaining the ability to measure variables and infer causation. Our week-long task had three between-subject conditions (to- tal N = 311): a CONTROL in which people engaged on Mag- pie with no external influence, and two (LEFT and RIGHT) in which a small number of posts were secretly made by us, shar- ing typical talking points from one political side. We found small but statistically reliable effects suggesting that, relative to the CONTROL, the presence of right-wing trolls resulted in a higher level of right-wing belief and a greater perception of po- litical division in the US. Conversely, the left-wing troll cam- paign did not appear to have any statistically reliable effect on these measures. We also found considerably more overall en- gagement in both troll conditions, probably because content with a clear political stance tended to receive more activity. However, participants (especially those on the left) disliked the RIGHT condition more than the others.

Prediction is central in human language processing, as the brain continuously predicts upcoming words using prior knowledge and context. Surprisal theory quantifies predictability using word surprisal. While previous studies link neural activity to surprisal during passive listening or reading, we investigate how surprisal is tracked in dynamic face-to-face conversations. Two key challenges arise: estimating surprisal as well as identifying predictions in EEG data in natural conversation. We address the first challenge by adapting a pre-trained large language model to a dataset of spontaneous conversation capturing features like hesitations and repetitions. We then relate the surprisal estimated by the adapted model to EEG data using temporal response functions. Our experimental results show neural tracking of surprisal at different time lags after word onset, supporting the surprisal theory in face-to-face conversation. To the best of our knowledge, we are the first to address the application of surprisal theory in such interactive settings.

Humans often interpret pointing as referring to an object, however, it can also indicate a direction or relevant spatial location. We investigated which one of these interpretations can explain 14-month-olds responses in a two-alternative choice task. We conducted three experiments, in which an experimenter pointed at one of the two lateral objects, swapped their positions in full view of the infant, and then allowed the infant to choose. Pointing was either produced in an Ostensive Addressing (Experiment 1), Nonostensive Addressing (Experiment 2), or Ostensive Labelling context (Experiment 3). In the Ostensive Addressing and Ostensive Labelling experiments infants chose the non-indicated object in the indicated direction significantly more often than predicted by chance. In contrast, in the Nonostenive Addressing experiment, infants' performance was on chance. These findings suggest that infants follow the direction of pointing rather than interpreting it as indicating a specific object in a communicative context.

The approximately number system (ANS) helps people quickly estimate the numerosity of objects in their environment. In this study, we explore one proposed mechanism for visually perceiving numerosity: visual clustering. Participants completed a magnitude comparison task, magnitude estimation task, and a clustering task using the same set of numerosity stimuli. The stimuli varied in the spatial configuration of the points (cluster structure -- clustered or dispersed) and in the number of points present. Participants judged stimuli with dispersed cluster structure to be more numerous in the magnitude comparison task. However, there was a minimal effect of cluster structure in the magnitude estimation task and no effect of the number of clusters perceived in both tasks. We also found that the clusters people perceived in the third task did not explain the effects of cluster structure. These findings go against strong claims that people use visual clustering to judge numerosity and set the stage for further investigations into the mechanisms underlying the ANS.

Proportional reasoning is essential for many real-world tasks, yet its developmental trajectory remains debated. Children's performance in nonsymbolic proportional reasoning varies across tasks and plummets when numerical information is misleading. The present study investigates whether 5- to 7-year-old children can accurately compare proportions in a naturalistic context where counting strategies are ineffective. Children listened to short stories in which a subset of people from each of two groups experienced an event (e.g., catching the flu). Given the equal numbers of affected individuals in both groups and different group sizes, children needed to rely on proportional reasoning to compare the prevalence of the event. Results showed that children performed significantly above chance overall. Moreover, they were more accurate in adverse scenarios (e.g., avoiding illness) than in favorable ones (e.g., acquiring rewards). These preliminary findings suggest that the ability to compare nonsymbolic proportions emerges by age 5 but varies depending on context.

Knowledge Graph Reasoning (KGR) is an effective way to ameliorate sparsity and incompleteness problems by inferring new knowledge based on existing knowledge. The probabilistic case-based reasoning (CBR) model can gather reasoning paths from similar entities and relations in KG, thus outperforming rule-based and embedding-based KGR methods. However, it is still limited by some problems, such as insufficient learning of similarity features and sparse intermediate activations. This paper proposes a \textbf{S}imilarity-\textbf{A}ware graph transformer-enhanced probabilistic CBR model for \textbf{KGR}, namely SA-KGR. The proposed model regards the reasoning task as the KG query answering and is composed of two phases. The first phase is similarity-aware graph transformer-based graph feature encoding, which equips the similarity matrix and Mixture-of-Expert network to obtain fine-grained similarity features that are more helpful for reasoning path generation. The second phase is similarity-enhanced probabilistic case-based reasoning, which can retrieve and infer query answers from the generated candidate paths to complete brain-like cognitive reasoning. Extensive experimental results on various benchmarks unambiguously demonstrate that the proposed SA-KGR model can obtain the state-of-the-art results of current CBR-based methods.

In the era of large-scale artificial intelligence, Large Language Models (LLMs) have made significant strides in natural language processing. However, they often lack transparency and generate unreliable outputs, raising concerns about their interpretability. To address this, the Chain of Thought (CoT) prompting method structures reasoning into step-by-step deductions. Yet, not all reasoning chains are valid, and errors can lead to unreliable conclusions. We propose ECCoT, an End-to-End Cognitive Chain of Thought Validation Framework, to evaluate and refine reasoning chains in LLMs. ECCoT integrates the Markov Random Field-Embedded Topic Model (MRF-ETM) for topic-aware CoT generation and Causal Sentence-BERT (CSBert) for causal reasoning alignment. By filtering ineffective chains using structured ordering statistics, ECCoT improves interpretability, reduces biases, and enhances the trustworthiness of LLM-based decision-making. Key contributions include the introduction of ECCoT, MRF-ETM for topic-driven CoT generation, and CSBert for causal reasoning enhancement.

Gaze Interaction Behavior Recognition (GIBR) plays a significant role in understanding social behaviors and diagnosing mental health conditions. However, existing methods are limited by inadequate task modeling, resulting in suboptimal performance. To address this issue, we model the GIBR task as a spatiotemporal reasoning problem integrating three modalities: emotion, gaze, and position. Based on this, we propose GIBRNet, which enhances the representation of gaze interaction tendencies through an Emotion-Aware Refinement Matrix and dynamically aggregates multi-frame, multi-modal information using GP-GNN, enabling more precise interaction behavior reasoning. Comparative experiments on the VACATION dataset demonstrate that GIBRNet significantly outperforms existing approaches. Additionally, we constructed a GIBR dataset suite, consisting of three extended datasets, for generalization evaluation, demonstrating GIBRNet's superiority. All datasets and code are publicly available.

This study investigates the Horizontal-Vertical Individualism-Collectivism (HVIC) traits of Large Language Models (LLMs), addressing the gap in understanding their cultural and social cognition. HVIC, a cross-cultural psychology framework, offers insights into cognitive patterns shaped by culture. We systematically evaluate multiple LLMs using quantitative (INDCOL scale) method, assessing their intrinsic HVIC traits and ability to simulate cultural and gender-based differences. Our findings reveal LLMs' capacity to capture HVIC nuances, providing a unique lens for studying human cognition through human-LLM comparisons. This research contributes to developing culturally sensitive AI systems and offers new perspectives on human HVIC traits, advancing both theoretical understanding and practical applications of AI.

The aim of this study is to analyse the auditory scene of musical works and to demonstrate that different compositions may prompt the emergence of distinctly interpreted perceptual streams in the listener's mind. The research focuses on selected excerpts from works by Alexandre Guilmant, Ludwig van Beethoven, and Antonio Vivaldi, which, due to their unique characteristics, elicit diverse auditory impressions. By combining score analysis with auditory scene analysis, this paper seeks to explain how different interpretations of the same sounds result in dissimilar auditory impressions. The auditory scene analysis presented here provides deeper insight into the process of stream formation and its implications for musical performance and aesthetic perception. The findings indicate that perceptual stream formation in music is considerably more complex and context-dependent than previously assumed, with implications for how listeners interpret auditory scenes.

Traditional education is instructor-centered. Providing exploratory learning activities before instruction typically engages students and improves learning. However, the design of exploratory learning activities can impact learning processes. This study investigated whether using tables or graphs in a statistics activity impacted exploratory learning processes and outcomes. Undergraduate students (N=252) in classroom and lab settings were taught about standard deviation. In instruct-first conditions, students received instruction, then an activity including either graphs or tables. In explore-first conditions, students explored either activity before instruction. After exploring, participants in the explore-first condition reported greater knowledge gaps and curiosity than the instruct-first condition. Graphical materials reduced cognitive load compared to tabular. However, instructional order and activity design did not impact learning outcomes (procedural knowledge, conceptual knowledge, representational transfer). Conceptual understanding was highest if students attempted multiple solutions while exploring graphical materials. Depth of exploration may affect conceptual benefits, especially when using graphical materials.

Exploratory learning before instruction has been effectively employed in STEM education to promote deeper conceptual understanding. However, its application to sociopolitical reasoning is underexplored. This research investigated whether exploratory learning can mitigate biased information processing, fostering more reflective evaluation of counter-attitudinal sociopolitical information. We examined how the order in which participants engaged with exploratory (data table) versus directly persuasive (verbal message) stimuli influenced the strength and confidence in sociopolitical beliefs. Participants reported increased support for positions they had initially opposed, and to large effect, regardless of stimuli order. However, an interaction of time and order on confidence levels hinted at potential metacognitive benefits for participants who explored first. Exploratory learning may be less beneficial in the context of sociopolitical decision making—at least when individuals are likely to update their beliefs anyway. However, exploratory learning might impact individuals' metacognition when the messenger contradicts their political position.

The rising prevalence of adolescent depression is a major public health concern. Current diagnostic methods, often burdensome and lacking objective biomarkers, hinder early detection. This paper proposes a novel multi-task framework integrating attention and resting EEG signals using Biosignal Learning and Agent Transformer (BLAT) for depression detection. EEG data is segmented, channeled, and positionally encoded, followed by feature extraction and classification via Agent Transformer. Experiments with 50 depressed adolescents and 50 controls achieved 85% accuracy using BLAT, offering a promising method for early adolescent depression screening.

We report here two behavioural experiments that investigated the Own Age Bias (OAB), better recognition performance of own-age vs. other-age faces, measured by the Face Inversion Effect (FIE). Both experiments employed an old/new recognition task where upright and inverted own (young adults) and other (older adults) where presented intermixed. Experiment 1 (n=48), used real-life faces, and revealed a robust OAB, where a significantly larger FIE (higher recognition for upright vs. inverted faces) was found for own vs. other age faces due to a reduced performance for upright other age faces. Experiment 2 (n=48) used standardised faces and revealed no effect of OAB, and no difference between upright own vs. other age faces. We interpret our results in the context of the perceptual learning and faces recognition literature.

Driver vigilance estimation plays a critical role in preventing fatigue-related traffic incidents. Current multimodal methods leveraging EEG and EOG signals often suffer from high computational costs due to the reliance on self-attention mechanisms like Transformers. To address these challenges, we propose a novel framework, Selective Attention Sequential Model (SAMM), which integrates a dynamic channel attention mechanism and the Mamba sequence modeling approach. By replacing traditional Transformer modules with Mamba's selective state spaces, our model achieves linear-time complexity while effectively capturing both local and global features.The SAMM framework fuses EEG and EOG signals using early fusion and employs a deep channel attention mechanism to enhance localized feature extraction. Mamba further complements this by efficiently modeling global dependencies in multimodal data, thus reducing computational costs while maintaining high accuracy. Extensive experiments on public datasets, SEED-VIG and SADT, demonstrate that SAMM achieves state-of-the-art performance with a significant reduction in inference time.

The proliferation of DeepFake has engendered widespread societal concerns, positioning its detection as a pressing imperative. Although existing studies have utilized single-subject EEG to distinguish between real and AI-generated content (AIGC), there is still a lack of research exploring dual-brain EEG and multimodal experimental paradigms. This study introduced a novel experimental paradigm, employing EEG hyperscanning to construct a dyadic EEG dataset for AIGC detection. This study employed inter-subject correlation (ISC) analysis to investigate the differences of interpersonal neural synchronization (INS). Additionally, this study proposed a novel neural network model named Squeeze-and-Excitation Depthwise Separable Convolution (SEDSC) for predicting the authenticity of real vs. AIGC. ISC analysis revealed apparent differences in INS under different modalities, valences, and animacy. Specifically, across the four frequency bands, both text and audio modalities elicited higher inter-brain synchronization under real materials than under AIGC materials. SEDSC utilized the phase locking value to assess inter-brain functional connectivity and weighted the inputs from four frequency bands before feeding them into the network for classification. This approach achieved a classification accuracy of 92.42% in distinguishing real from fake content. This study designed a new experimental paradigm and constructed a dataset, confirming that there were evident differences in INS during tasks involving real and AIGC materials. Furthermore, SEDSC successfully predicted the authenticity of the content.

This paper examines the role of cognitive biases in the decision-making processes of large language models (LLMs), challenging the conventional goal of eliminating all biases. When properly balanced, we show that certain cognitive biases can enhance decision-making efficiency through rational deviations and heuristic shortcuts. By introducing heuristic moderation and an abstention option, which allows LLMs to withhold responses when uncertain, we reduce error rates, improve decision accuracy, and optimize decision rates. Using the Balance Rigor and Utility (BRU) dataset, developed through expert collaboration, our findings demonstrate that targeted inspection of cognitive biases aligns LLM decisions more closely with human reasoning, enhancing reliability and suggesting strategies for future improvements. This approach offers a novel way to leverage cognitive biases to improve the practical utility of LLMs across various applications.

To understand language, we use knowledge about everyday events to create rich internal (situation) models. Although knowledge increases with age, fluid cognitive abilities tend to decline, potentially making it more difficult to access that knowledge. Here, we asked how aging affects the ability to use event knowledge during real-time language comprehension. We recorded event-related brain potentials as younger and older adults read vignettes about everyday events. Both groups showed facilitation on the N400 (a neuroelectric marker of semantic processing) for words that fit the context. However, only younger adults showed facilitated N400s to anomalous but event-related words compared to unrelated anomalies. Among older adults (aged 53-80), there was a negative correlation between age and N400 effects of event-relatedness. We conclude that real-time access to event knowledge during language comprehension may shift across the course of the adult lifespan such that older adults restrict activation to the most immediately relevant content.

Assessments of early cognitive and linguistic abilities typically involve picture stimuli. As these assessments spread worldwide, researchers make an implicit assumption: that children across contexts understand pictures in the same way, at the same developmental timepoint. What if this assumption does not hold for some or all kinds of pictures? In the present research, a preregistered sample of 128 3- to 7-year-olds from Kisumu, Kenya participated in a Swahili vocabulary assessment. Using a within-subjects design, each participant completed vocabulary trials in four formats (i.e., objects, photographs, cartoons, black-and-white line drawings). Preregistered analyses showed that children performed equally accurately across object, photograph, and cartoon formats, but less accurately in the line drawing format. However, exploratory analyses suggested that a subset of line drawings drove this difference. These findings suggest that caution is necessary in the use of picture stimuli and that assessments involving line drawings may sometimes underestimate children's capacities.

This study explores the role of coherence-based reasoning in belief updating within uncertain environments. We develop a novel computational model where agents update their beliefs based on observed evidence, with some evaluating the coherence of their belief set before accepting new evidence. Our results show that coherence-based evidence filtering improves belief accuracy in noisy environments and when agents' prior beliefs are accurate. However, when agents encounter systematically misleading evidence, coherence considerations lead to less accurate beliefs. These findings shed light on how coherence interacts with evidence quality and belief accuracy.

The role of mental simulation in human physical reasoning is widely acknowledged, but whether it is employed across scenarios with varying simulation costs and where its boundary lies remains unclear. Using a pouring-marble task, our human study revealed two distinct error patterns when predicting pouring angles, differentiated by simulation time. While mental simulation accurately captured human judgments in simpler scenarios, a linear heuristic model better matched human predictions when simulation time exceeded a certain boundary. Motivated by these observations, we propose a dual-process framework, Simulation-Heuristics Model (SHM), where intuitive physics employs simulation for short-time simulation but switches to heuristics when simulation becomes costly. By integrating computational methods previously viewed as separate into a unified model, SHM quantitatively captures their switching mechanism. The SHM aligns more precisely with human behavior and demonstrates consistent predictive performance across diverse scenarios, advancing our understanding of the adaptive nature of intuitive physical reasoning.

Large Language Models (LLMs) excel in understanding and generating text but struggle with providing professional literary criticism for works with profound thoughts and complex narratives. This paper proposes GLASS (Greimas Literary Analysis via Semiotic Square), a structured analytical framework based on Greimas Semiotic Square (GSS), to enhance LLMs' ability to conduct in-depth literary analysis. GLASS facilitates the rapid dissection of narrative structures and deep meanings in narrative works. We propose the first dataset for GSS-based literary criticism, featuring detailed analyses of 48 works. Then we propose quantitative metrics for GSS-based literary criticism using the LLM-as-a-judge paradigm. Our framework's results, compared with expert criticism across multiple works and LLMs, show high performance. Finally, we applied GLASS to 39 classic works, producing original and high-quality analyses that address existing research gaps. This research provides an AI-based tool for literary research and education, offering insights into the cognitive mechanisms underlying literary engagement.

Why do large-scale field experiments in education often have muted effects? Drawing on system dynamics and self-regulated learning theory, we sought answer this question by simulating the behavior of self-regulated (discrepancy-reducing) learners over time affected by different types of educational interventions. We analyze three types of interventions: changing students' learning rates (learning strategies), intercepts (prior knowledge or teaching effectiveness), and norms of study (achievement goals). We uncover situations where educational interventions can affect achievement in the short run, but typical cross-sectional analyses do not find a measurable effect in the long run. Results indicate that highly motivated, self-regulated learners may resist external interventions, particularly those targeting learning strategies or prior knowledge. In contrast, interventions show the greatest effect on achievement when students are under time constraints and struggling to achieve their desired performance. Ultimately, self-regulated learners may be the hardest to help, a phenomenon we call the "self-regulated learning paradox."

This study investigates how Large Language Models (LLMs), specifically Baidu's Ernie Bot, shape personal narratives when users seek emotional support over repeated sessions. Sixteen participants from China engaged in weekly chatbot interactions for four weeks, supplemented by reflective diaries and pre-/post-study interviews. Conversation analysis and quantitative measures (e.g., mood ratings, meaning-making scales) revealed incremental shifts in user language, including increased lexical alignment with AI-generated phrases and more positive emotion words. In-depth interviews highlighted the complex process by which participants alternately embraced or resisted the AI's framing, with many reporting newfound perspectives and a sense of empathic resonance. However, some voiced skepticism regarding the AI's genuine capacity for emotional understanding, underscoring ethical dilemmas related to anthropomorphism and data privacy. Overall, the findings suggest that iterative dialogues with an empathic-seeming LLM can facilitate meaningful narrative reframing, albeit with notable variations in user experience and potential risks of over-reliance.

Emotion recognition using electroencephalography (EEG) represents a significant area of study in brain-machine interfaces. To address this multifaceted challenge, it is crucial to improve the ability of EEG features to represent emotional states. A hypergraph-based methodology allows for the depiction of higher-order spatial correlations to develop distinguishing emotional features. However, the original hypergraph may lack robustness due to potential interference among local channels. In addition, excessively coarse hypergraph granularity can result in the loss of critical information. To mitigate these issues, we propose hypergraph group learning, which aims to balance robustness with the retention of detailed information. In this study, we model temporal and spatial dependencies across varying granularities using Hypergraph Group Learning to achieve a discriminative representation of emotional features. We used multiple CNN convolutions to map EEG signals from different brain regions and time segments into a unified distribution. The multi-granularity hypergraph convolutional network (MGHGCN) is specifically designed to capture long-term temporal correlations among channels effectively. By integrating multiview fusion, we significantly improved the accuracy and robustness of EEG-based emotion recognition. Experimental results from publicly available datasets, including SEED, SEED-IV, and EMOT, validate the effectiveness of our approach, achieving precisions of 98.51 (2.46) %, 89.20 (6.13) % and 97.79 (1.31) %, respectively. These results demonstrate that our hypergraph effectively maintains both robustness and detailed information.

Cognitive impairment is a common disease among middle-aged and elderly people, which seriously affects health outcomes and quality of life, and carries a risk of progressing to severe stages such as dementia. Early identification is beneficial for timely intervention and treatment. This study proposes a new model for predicting cognitive impairment that integrates static and dynamic data, including medical, demographic, and social relationship features. It combines Granger causality with deep learning and uses multiple metrics to evaluate model performance. The performance comparison results between our model and the baseline model demonstrate that our model's predictions have a certain level of accuracy. In addition, causal features derived from Granger causality analysis are used to identify cognitive impairments. Statistical analysis shows that the selected features have statistical significance, further verifying the robustness of our model and its potential for predicting cognitive impairment.

The interaction between hippocampal ripples and cortical slow oscillations (SO) has been proposed to play a critical role in memory consolidation during sleep. However, the neuronal mechanisms underlying the transmission of ripples within cortical regions remain poorly understood. In this study, we used a computational model to investigate how ripple events propagate through cortical networks. We found that sparse and weak inter-areal connections impede ripple propagation, while dense and strong inter-areal connections facilitate it. Notably, our findings reveal that when cortical networks exhibit slow oscillations (SOs), characterized by alternating up and down states, ripples occurring before the SO peak can propagate to distant cortical areas even in the presence of sparse and weak inter-areal connections. These results indicate that the precise coordination between SO and the ripple promotes efficient communication in cortical regions during sleep. This study offers new mechanistic insights into the role of SOs during slow-wave sleep, deepening our understanding of the processes underlying memory consolidation.

In this work, we outline a methodological analogy between cognitive sciences and physics regarding the use of models and the synthetic method. Beginning with brief historical remarks on ‘the discovery of the artificial' (as defined by Roberto Cordeschi) in early 20th century behavioral sciences and the methodological turning point in statistical physics at the same time, we demonstrate that the ‘envy' for the use of analytical theories in the exact sciences—often referred to as physics envy—which has significantly influenced the development of psychology and the sciences of human behavior, is ultimately unfounded. Finally, we use this ‘overcoming' of the physics envy, along with some brief considerations on notable 20th century theoretical results related to ‘limitation' of computability and complexity, to demonstrate how cognitive sciences —like the natural sciences— must necessarily rely on models and simulations. This necessity arises from the inherent complexity of the systems under study, which precludes their treatment in analytical and exact terms.

Similar Place Avoidance (SPA) is the cross-linguistic tendency whereby languages avoid transvocalic consonants with the same place of articulation within a word. In this study, we examine if SPA is the result of learning biases against words where the consonants share a place of articulation. In two experiments we examine whether adults show a learning difference between place-disagreeing novel words (e.g. [tip]) and place-agreeing novel words (e.g. [tid], where [t] and [d] are coronal). Participants are taught novel words and are then tested in an object-mapping or lexical decision task. We measure participants' learning performance based on accuracy and reaction times. Results indicate that, while accuracy is comparable for place-agreeing and place-disagreeing words in both tasks, participants' lexical decision responses are generally slower for place-agreeing words. These results suggest that participants experience processing difficulties when accessing newly-formed representations of place-agreeing words, which may contribute to the existence of SPA.

Face processing is widely regarded in cognitive science as the integration of individual features into a holistic percept. However, recent neuroscience research highlights a more nuanced interplay between holistic and featural mechanisms, with specific facial features receiving greater emphasis during early perception. Event-related potential studies reveal that the number and type of parafoveal features significantly influence neural response delays, yet the underlying mechanistic model remains unclear. This paper examines these phenomena through the lens of the predictive coding network, a biologically plausible alternative to traditional deep neural networks. Our findings show that predictive coding networks accurately simulate the influence of parafoveal features on neural response times while upholding the saliency hierarchy of facial features. These results provide a computational explanation for the observed neural delays and highlight the potential of predictive coding as a robust framework for understanding face perception in the human brain.

Multilingual large language models (MLLMs) typically underperform when answering questions in non-native languages compared to their native language. Although existing translate-then-answer prompting methods partially alleviate this issue, their performance remains suboptimal compared to directly answering questions in the native language. Moreover, current studies lack a clear explanation for this gap. In this study, we attribute this issue to incomplete Cognitive Priming, a phenomenon observed from human cognition. And we point out that while existing methods achieve Language Priming (LP), they overlook Domain Priming (DP). To address this, we propose Cognitive Priming Prompting (CogPrim), which employs a Role-Enhanced Multi-MLLM Collaboration strategy to ensure both LP and DP, thereby improving knowledge elicitation for non-native QA tasks. Across five language question-answering benchmarks, CogPrim outperforms all state-of-the-art related methods. This approach contributes to advancing the understanding of human-like cognitive behaviors in MLLMs, fostering better MLLM service to a broader range of multilingual users.

This study investigates how integrating human eye-tracking data into Large Language Models (LLMs) and Visual Large Language Models (VLLMs) can enhance document comprehension in tasks that require both linguistic understanding and visual attention, specifically Semantic Entity Recognition (SER) and Document Question Answering (DQA). Despite rapid advancements in AI-based document understanding, LLMs still face challenges in replicating the depth of human cognition, particularly in how reading order and visual attention affect comprehension. The results demonstrate that human reading order and the regions they focus on significantly impact performance in both tasks. Additionally, while LLMs do not need to fully mimic human reading sequences, their performance improves when their attention patterns align more closely with human visual strategies. This highlights the importance of incorporating cognitive-inspired attention mechanisms in AI systems, offering a path to better AI models that reflect human cognitive strategies in complex document understanding.

The present study employed a creativity-on-demand task to investigate the variability of discourse biases associated with Implicit Causality verbs when language users are demanded to come up with original discourse continuations. While Implicit Causality turned out to be unaffected by the creativity manipulation, the likelihood to continue with a contiguity relation was influenced by the creativity manipulation. These findings are in line with the Two-Mechanism Account by Solstad & Bott (2022, 2023) grounding Implicit Causality in semantic constraints imposed by lexical semantics and relating other next-mention biases to the Contiguity Principle, a more general discourse pragmatic principle.

Research has indicated that baseline pupil size may reflect attentional control capabilities. Several studies have demonstrated links between resting pupil measurements and various cognitive functions, including working memory capacity and fluid intelligence. The existing literature presents mixed evidence, with some studies supporting this relationship while others fail to replicate these findings. Moreover, the validity of this relationship across different populations has come into question, particularly when considering age-related physiological changes that affect pupil size. Our study sought to investigate whether the relationship between baseline pupil size and attentional control remains consistent across age groups. Results showed no meaningful correlation between baseline pupil size and attentional control. In examining possible explanations for these disparate results, we identified several methodological challenges, including inconsistencies in testing environments and variations in pupil measurement protocols, that may account for the conflicting findings in previous research. These observations highlight the importance of developing more standardized experimental approaches to properly evaluate baseline pupil size as an indicator of attentional control, particularly when studying aging populations.

This paper presents a big data urban visualization platform for Guangdong Province, aimed at enhancing the efficiency and intelligence of urban planning. The platform utilizes Python to collect city-specific data, with data storage implemented using a MySQL database, complemented by NoSQL technologies to support the integration of unstructured data. The Flask backend employs deep learning and data mining algorithms to identify complex relationships among urban data, and we have also integrated graph neural network methods to capture spatial dependencies across different geographic regions within the city. Meanwhile, the ECharts frontend generates dynamic charts to present diverse information. Through a front-end and back-end separation architecture, the system ensures real-time updates, enhancing user experience. This research further emphasizes the need to explore challenges in real-time data integration, expanding data sources, optimizing user interactions, and protecting data privacy, providing important directions for future AI-driven urban planning.

Personalized recommendation aims to recommend candidate items to users based on their preferences by simulating their cognitive decision-making process. User-item interaction data typically follows a power-law distribution. However, existing works usually learn the representations of users and items in Euclidean space, resulting in a mismatch between the data volume space and the embedding space, which causes significant distortion in the representations. Moreover, the presence of cognitive biases, such as conformity, can also introduce distortion in representation learning. Therefore, we propose a Debiased Hyperbolic Recommendation model, called DHRec. Specifically, first, we choose to model the representations of user and item in hyperbolic space, which has exponential growth capabilities. Second, in addition to the user-item interaction graph, we also construct semantic graphs to capture the semantic neighbor information of users and items. Then, by adjusting the weights of neighbor nodes, we learn debiased representations of users and items, effectively alleviating the bias caused by conformity. Finally, we compute the predicted scores between user and candidate items in hyperbolic space. Extensive experiments on three datasets demonstrate that our model surpasses the strongest baseline, achieving a 11.04% and 10.09% improvement on Recall and NDCG, respectively.

Parkinson's disease (PD), characterized by its incurable nature, rapid progression, and severe disability, poses significant challenges to the lives of patients and their families. Given the aging population, the need for early detection of PD is increasing. In vitro diagnosis has garnered attention due to its non-invasive nature and low cost. However, existing methods present several challenges: 1) limited training data for facial expression diagnosis; 2) specialized equipment and acquisition environments required for gait diagnosis, resulting in poor generalizability; 3) the risk of misdiagnosis or missed diagnosis when relying on a single modality. To address these issues, we propose a novel multimodal in vitro diagnostic method for PD, leveraging facial expressions and behavioral gait. Our method employs a lightweight deep learning model for feature extraction and fusion, aimed at improving diagnostic accuracy and facilitating deployment on mobile devices. Furthermore, we have established the largest multimodal PD dataset in collaboration with a hospital and conducted extensive experiments to validate the effectiveness of our proposed method.

In recent years, the remarkable performance of large language models (LLMs) in tasks such as legal judgment prediction (LJP) has garnered widespread attention. An increasing number of LLMs have been successfully implemented to assist judges in performing various legal tasks. However, their robustness and reliability in complex judicial scenarios remain a subject of debate, particularly when confronted with real-world legal cases. Existing research often overlooks the systematic evaluation of these LLMs in terms of judicial fairness, robustness and other ethical considerations. To fill this gap, we propose a novel benchmark that integrates authentic legal cases to evaluate the robustness of LLMs in the legal judgment prediction (LJP) task. Our work establishes foundational safety standards for applying LLMs in the legal domain.

Relational reasoning develops slowly. Children's difficulty may stem from the difficulty of inhibiting object attention, but the role of object attention in relational reasoning remains unclear. Experiment 1 tested the hypothesized trade-off between processing relational and object information using a novel match-then-recognize paradigm. The relational-mapping task required participants to match ordinal positions of objects, followed by an object memory test. In adults, there was a clear trade-off: object recognition was negatively correlated with relational mapping. Children's object recognition was lower than adults' but positively correlated with relational matching. Experiment 2 further tested the role of inhibition by occluding object matches. Surprisingly, older children and adults actively removed occluders, to their own detriment, whereas the youngest children neither removed occluders nor were helped by them. Findings suggest that selective attention can be crucial for relational reasoning, but ignoring object information may be less important for success than total allocated attention.

Motivated by the two contrasting forces in shaping linguistic knowledge—efficiency and redundancy, the present study examines sentence comprehension behaviour amongst Korean children aged three to six, focusing on verbs (relative to case markers) in interpreting transitive events. Through picture-selection experiments that systematically omit and obscure portions of transitive sentences, we find (1) a reduced role of verbs in sentence comprehension and (2) age-related variations in the application of case-marking knowledge. These findings suggest that verbs may sometimes become redundant during comprehension, which is attributable to early maturation and strong automatisation of verbs. This provides support for verb-periphery strategies so as to maximise efficiency in language activities amongst Korean preschoolers.

Building on Shin (2022), the present study examines how Korean monolingual children comprehend suffixal passive constructions by employing a webcam eye-tracking method, aiming to test two theoretical accounts of grammatical generalisation (gradual vs. early abstraction). Twenty-eight children aged three to six, alongside 20 adults, joined picture-selection experiments paired with eye-gaze measurements. The findings indicate that children's utilisation of passive-voice heuristics remains limited yet developing, overshadowed by well-entrenched active-voice knowledge. In particular, the eye-gaze data reveal processing challenges related to the passive voice, mainly interpretive difficulties arising from passive morphology. These results replicate those of Shin (2022), offering further support for a moderate version of each account that emphasises the pivotal role of linguistic exposure in mastering linguistic knowledge. From a methodological standpoint, this study enhances the accessibility of webcam eye-tracking research for understudied languages in the field.

Food choices, as a type of value-based decision, are affected by environmental cues. We conducted three studies to investigate the computational mechanisms through which background music influences the food choice-making process. Hierarchical drift-diffusion modeling revealed that nature-related music led to higher drift rates than urban-related music, indicating faster evidence accumulation toward certain choices. Specifically, participants processed the value of vegetable-forward meals more efficiently when exposed to nature-related music compared to urban-related music. Moreover, the effect of music on the drift rate varied with the vividness of music-induced mental imagery and the perceived identity of performers (human or robot). Collectively, these findings reveal the computational mechanisms underlying the influence of environmental cues on value-based decision-making.

Human adults think of the natural world as orchestrated by rules, yet many of them are neither equally rigid nor clearly evident. Some are beset by exceptions, and others are not intuitive. The problem of rule learning is especially salient in development, as children are continuously learning how the world works. What cognitive mechanisms underpin this rule learning? We propose a computational model for formulating and testing hypotheses in naturalistic contexts, that combines Bayesian inference under uncertainty over self-generated and social evidence with formal rules and optimistic information-seeking heuristics. We validate our model experimentally, showing that it explains 7- to 10-year-olds' behavior in a rule-based, physical task, including the distribution and the types of evidence children sampled. The proposed model outperforms both a purely rule-based Bayesian hypothesis search and a resource-rational random sampling approach. Our results suggest that children implement an internal mechanism for generating and testing a limited number of hypotheses, including formal programmatic rules and heuristics generated from salient problem features to seek more evidence when formal rule generation fails.

With the growth of the metaverse and XR technologies, BCIs are expanding from medical applications to various consumer industries. However, this broader adoption has raised concerns about the privacy and security of EEG data. To address this challenge, we propose the RPW-EEG framework, which embeds copyright information as perturbations to enhance the security and traceability of data, while ensuring the robustness and usability of the data. The framework adopts an encoder-decoder architecture for end-to-end training, incorporating a noise layer to enhance the stability and anti-attack capabilities of the watermark data. Additionally, to prevent the loss of task-related features in EEG data, we introduce a plug-and-play fine-tuning module that restores these features within the watermark-embedded signals. Experimental results show that RPW-EEG outperforms baseline models in watermark quality, with extraction accuracy exceeding 61% under various attacks, and achieves 88.5% classification accuracy for task paradigms, effectively balancing copyright protection and EEG data usability.

Automated vehicle (AV) validation faces the "billions of miles" challenge, requiring high-fidelity simulations to replicate diverse interactive driving behaviors for safety. Traditional methods oversimplify by using uniform behavioral models, ignoring the diversity of human driving styles, which are deeply influenced by individual psychological traits. This research introduces a condensed framework for representing interactive driving styles, by incorporating these psychological dimensions, balancing completeness and complexity. Key features include: i) individual style recognition via attention mechanisms and hierarchical contrastive learning, capturing subtle cognitive-based interaction patterns that reflect underlying differences in driver psychology (e.g., risk tolerance, decision-making heuristics); ii) scenario-independent style compression, filtering external factors to extract intrinsic driver intentions; iii) dimensionality-aware refinement, mapping complex behaviors to low-dimensional psychological axes for efficient computation. Tests on the NGSIM dataset reduced testing complexity by decoupling styles from scenarios. Compared to traditional methods, style distinctiveness improves by 28% (entropy-based), with 85% edge-case behavior coverage. This framework supports scalable AV testing by integrating diverse, psychologically-informed driving styles without combinatorial complexity.

Understanding how others perceive the tactile world is essential for human communication. Each individual has a unique and multidimensional tactile perceptual space, which makes it challenging to understand others' tactile perceptions within the spaces. This study investigates the cognitive ability and process of estimating another person's subjective similarity of various textures, a key aspect of tactile perceptual space. Participants performed tasks to estimate how a target individual would rate the similarity of tactile stimuli pairs. Results showed that participants could partially infer target's similarity ratings, and their estimated tactile similarity ratings converged toward an intermediate point between their own and the target's. The estimation process included exploration, adaptation, and overgeneralization phases. Moreover, participants' own similarity ratings shifted closer to those of the target after performing the task. These findings suggest that estimating another person's subjective similarity between textures involves incorporating elements of the target's tactile perceptual space into one's own.

Making sense of the world and acting in it relies on building simplified mental representations that abstract away aspects of reality. This principle of cognitive mapping is universal to agents with limited resources. Living organisms, people, and algorithms all face the problem of forming functional representations of their world under various computing constraints. In this work, we explore the hypothesis that human resource-efficient planning may arise from representing the world as predictably structured. Building on the metaphor of concepts as programs, we propose that cognitive maps can take the form of generative programs that exploit predictability and redundancy, in contrast to directly encoding spatial layouts. We use a behavioral experiment to show that people who navigate in structured spaces rely on modular planning strategies that align with programmatic map representations. We describe a computational model that predicts human behavior in a variety of structured scenarios. This model infers a small distribution over possible programmatic cognitive maps conditioned on human prior knowledge of the world, and uses this distribution to generate resource-efficient plans. Our models leverages a Large Language Model as an embedding of human priors, implicitly learned through training on a vast corpus of human data. Our model demonstrates improved computational efficiency, requires drastically less memory, and outperforms unstructured planning algorithms with cognitive constraints at predicting human behavior, suggesting that human planning strategies rely on programmatic cognitive maps.

Previous work suggests that non-visual object properties like weight are automatically integrated into event models during language comprehension. Horchak and Garrido (2021) found that Portuguese speakers' response times were faster when the state of a presented object (e.g., a smashed tomato) matched the event implied by the preceding sentence (e.g., You drop a bowling ball on a tomato). In an exact replication in English (Experiment 1), we failed to replicate this weight-state match effect. In Experiment 2, we examined the potential role of sentence focus, manipulating whether the target item served as the subject or direct object in the sentence. Response times revealed a weight-state match effect, but only when the target object was the focus (i.e., subject) of the sentence. Overall, these findings suggest that the representation of object state changes during language comprehension may depend on the interaction of object properties and language-specific syntactic constraints.

Lexical networks may vary as a function of individual differences in vocabulary knowledge and word-level features. Analyses often rely on descriptive network statistics, which do not support robust inferences. This study introduces the latent space model as a method for assessing the degree to which network structure is accounted for by word-level features. We analyze lexical networks from adults with below-average vs. above-average receptive vocabulary knowledge (n = 22 per group). We used latent space models to assess effects of semantic, taxonomic, and phonological similarity between words on network structure as well as effects of part-of-speech, concreteness, age-of-acquisition, and word frequency. For both groups, we found significant effects of semantic and taxonomic similarity, with additional effects of phonological similarity and concreteness for the low vocabulary group. These findings suggest increased reliance on fewer cues in lexical networks of adults with larger vocabularies. Implications for inferential modeling of lexical networks are discussed.

Thurstone models have been widely applied in wisdom-ofthe-crowd applications to aggregate individual rankings due to their ability to represent individual knowledge and achieve high accuracy. However, they lack the ability to generalize even across highly similar items and cannot leverage external knowledge bases or learned machine representations. In this work, we extend Thurstone models for partial ranking data by introducing a latent construct that maps pretrained vector representations to latent truths. These representations are finetuned through a single neural network layer, enhancing the model's ability to capture meaningful ranking structures. We evaluate our neural Thurstone model across objective ranking tasks, including animal speeds, material hardness, and the longitudinal positioning of U.S. states from west to east. Our results demonstrate that the extended model improves aggregation accuracy in sparse data settings and generalizes to novel items with moderate predictive accuracy, highlighting its potential to enhance collective intelligence in ranking-based inference.

Word meaning extension refers to the process by which a single word form develops multiple related meanings. Young children exhibit the capacity to extend word meaning, and previous research shows that such word overextension relies on multimodal semantic knowledge. We explore the evolutionary trace of word meaning extension by asking whether nonhuman animals might have this shared capacity. We compare meaning extension in children with the attested cases of overextension collected from the YouTube channel of a Grey Parrot, Apollo, who has acquired some English words. Our results show that parrot overextension can be predicted by a multimodal child overextension model better than baselines, which demonstrates that Grey Parrot may be using semantic knowledge similar to children for choosing words to express new referents. Our finding suggests that meaning extension is a cognitive ability identifiable in species about 320 million years apart from humans.

Limited work has considered how the temporal distribution of labeling events affects word learning in ambiguous contexts, such as the cross-situational word learning paradigm, over real-world timescales. The temporal distribution of learning events can impact how well information is retained: spacing out information promotes retention more than presenting information in close succession. In the current study, adults were presented with novel object-word pairings across six different temporal schedules over four consecutive training days. Word learning was assessed either immediately after the final training session (N = 50) or after a week (N = 54). Results revealed that adults successfully disambiguated word-object mappings across all learning schedules at both test times, except for the massed and most spaced schedules at the 1-week delay. These findings suggest that temporal distribution effects emerge across extended timescales, but there might be constraints on the amount of spacing that is optimal for word learning.

The Rational Speech Act (RSA) model has been employed to explain word learning and inferences based on the costs of forms. Here, we focus on a hitherto untested and counterintuitive cost-based effect predicted by RSA: In learning a lexicon with two forms and meanings, learners should prefer an ambiguous costly form over an ambiguous cheap form. We demonstrate this prediction in an RSA model. We then measure reaction times and lexicon ratings in a novel word learning task to test whether a lexicon with an ambiguous costly form is less surprising than one with an ambiguous cheap form. We found no clear evidence for this effect in either measure. We discuss alternative explanations for documented cost-based effects, and the possibility that cost-based inferences may not occur during learning.

To tackle the issue of hallucination in generative question answering (GQA)—where the generated answer is nonsensical or unfaithful to the provided document—we introduce a novel framework called evidence-enhanced triplet generation (EATQA). This framework incentivizes the model to generate all possible combinations of ⟨Question, Evidence, Answer⟩ triplets by reversing the source pair and target label to grasp their logical interrelationships. Specifically, the model predicts the Answer (A), Question (Q), and Evidence (E) given the QE, EA, and QA pairs, respectively. Furthermore, we address the distribution gap during the inference stage to extract knowledge from the evidence more effectively. Our framework ensures that the model comprehends the logical connections between queries, evidence, and answers, thereby simultaneously enhancing evidence generation and question answering capabilities. In this study, we apply the EATQA framework to the LLama model, demonstrating superior performance compared to other large language model (LLM)-based methods and hallucination mitigation techniques on two challenging GQA benchmarks. Further analysis reveals that our method not only preserves the pre-existing knowledge within the LLM but also reduces hallucination and produces more accurate answers.

Empathy has been linked to music preferences in controlled laboratory settings, but naturalistic settings like music streaming platforms remain unexplored. This study investigates how trait empathy influences music preferences and sequential listening behaviors in real-world settings. To this end, we collected and analyzed one-year Spotify listening histories of 290 Indian university students alongside their trait empathy scores, measured using the IRI scale. Our results reveal that individuals who score high on the IRI subscales of Empathic Concern, Fantasy, and Perspective Taking prefer sad music. Moreover, those scoring high on Empathic Concern or Perspective Taking were found to be more likely to transition from happy to sad music. These findings partially align with previous lab-based research, specifically for the subscales of Empathic Concern and Fantasy, while providing novel insights into the relations between the Perspective Taking subscale and music consumption. The study also provides novel insights into sequential listening behaviours, thus, strengthening the evidence that empathy shapes musical preferences and listening behaviors across diverse contexts.

We adapted Irving's (2014) Image Control and Recognition Task (ICRT) to explore a phenomenon we term outsight. The ICRT is a visual synthesis task: Participants construct a mental image of an object following stepwise instructions. They are then asked to name and subsequently draw the imagined object. We focus on trials when participants after having failed to name their mental image, could do so after having drawn it. In this exploratory study, such outsight recognition occurred on 29% of the ICRT trials. In addition, outsight recognition was accompanied by some of the phenomenological markers associated with aha! experiences. We offer some reflections on the importance of reified imagery for creativity.

The challenge of early word learning is often framed as one of individually mapping words to their referents. Yet children do not experience words just as individual labels, but as parts of broader language contexts, such as conversations and stories. In principle, word contexts might support word learning because words similar in meaning tend to occur in similar contexts. Thus, a child who knows some fruit words and has heard them in the context of "juicy" might learn that a "juicy mango" is also a fruit even without ever seeing a mango. Although children can use such contextual support to learn words in the lab, we do not know whether they harness contextual support in everyday language for real-world word learning. We quantified words' contextual support in children's everyday language input and found that it reliably predicted normative word learning, even accounting for other established predictors such as word frequency.

When learning with self-testing alone, will a learner make inferences between the tested items? This study examines whether self-testing's benefits extend beyond isolated facts to support broader connections between the facts. Comparing self-testing to self-explanation (a strategy known to facilitate inferential learning), we find that while self-testing participants show superior recall of individual facts, they perform significantly worse at making connections between those facts.

In language comprehension, mental representations of temporal relations between described situations are often construed by inference. While the basis for these inferences remains unclear, growing evidence suggests that abstract predicate properties – such as dynamicity and causal structure – play a crucial role in temporal event construal. Across two self-paced reading experiments, replicating and extending Gennari (2004), we find that temporal proximity inferences are shaped by these properties, but only for stative predicates that generally encode non-dynamic situations without causal structure: Participants consistently expected states to overlap in time, in both complement (Experiment 1) and relative clause constructions (Experiment 2). These findings indicate that temporal proximity inferences arise as a general feature of (non)-dynamicity, supporting models of language comprehension that prioritize abstract event structural properties in shaping temporal inferences.