Official implementation of the paper Hallucination Detection in LLMs Using Spectral Features of Attention Maps, accepted at EMNLP 2025 (see how to cite our work).

• Python 3.12+
• uv package manager

Install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

CPU (Linux/macOS):

make install_cpu

GPU (Linux with CUDA 12.4):

make install_gpu

The following flowchart describes the main steps to reproduce experiments. All steps can be run using DVC stages defined in dvc.yaml. In addition, dvc.yaml define more stages to compute results for ablation study. Below, we describe the main steps in more detail.

graph LR
    A[Generate Attention Diagonals & Answers] --> B[Generate Labels]
    B --> C[Generate Split]
    C --> D["Train LapEigvals/AttnEigvals/AttnLogDet"]
    C --> E["Compute AttnScore (LLMCheck) baseline"]
    F["Generate Hidden States"] --> G["Train Hidden States Baselines"]
    C -->|"Re-use labels from attention features"| F

• CoQA - download devset from the official website: https://nlp.stanford.edu/data/coqa/coqa-dev-v1.0.json
• GSM8K - is available through huggingface hub, will be downloaded automatically.
• HaluevalQA - download data from the official repository: https://github.com/RUCAIBox/HaluEval?tab=readme-ov-file#data-release
• NQOpen - is available through huggingface hub, will be downloaded automatically.
• SQuADv2 - download devset from the official website: https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json
• TriviaQA - is available through huggingface hub, will be downloaded automatically.
• TruthfulQA - is available through huggingface hub, will be downloaded automatically.

1. Generate attention diagonals and answers:

CUDA_VISIBLE_DEVICES=0 NUM_PROC=1 dvc repro generate_attentions_only

CUDA_VISIBLE_DEVICES=0 NUM_PROC=1 dvc repro generate_hidden_states_for_selected_tokens

2. Evaluate generated answers

dvc repro eval_answers_ngram

3. Evaluate generated answers using LLM-as-judge

dvc repro eval_answers_llm_judge

4. Generate labels

dvc repro generate_labels

dvc repro generate_labels_gsm8k

5. Generate split

dvc repro generate_split

6. Train probes

dvc repro train_attn_vs_laplacian_pca

dvc repro train_hidden_states_baselines

dvc repro probe_attn_score

If you use this code in your research or find the work relevant, please consider citing our paper:

@inproceedings{binkowski2025hallucination,
  title={Hallucination Detection in {LLM}s Using Spectral Features of Attention Maps},
  author={Jakub Binkowski and Denis Janiak and Albert Sawczyn and Bogdan Gabrys and Tomasz Jan Kajdanowicz},
  booktitle={The 2025 Conference on Empirical Methods in Natural Language Processing},
  year={2025},
  url={https://openreview.net/forum?id=tm5JQTpBhj}
}