Research map / Large language models / Language, safety and interpretability
Detecting Contextual Hallucinations in Large Language Models with Frequency-Aware Attention
Siya Qi, Yudong Chen, Runcong Zhao, Qinglin Zhu, Zhanghao Hu, Wei Liu, Yulan He, Zheng Yuan, Lin Gui
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| extract high-frequency attention components instead of using full-spectrum or low-pass features (ablation) | Fourier attention features · RAGTruth and HalluRAG (avg AUROC) | AUROC | Improved |
| use low-pass or full-spectrum attention features instead of high-pass (ablation) | Fourier attention features · RAGTruth and HalluRAG (avg AUROC) | AUROC | No clear effect |
| replace attention variance/entropy summaries with frequency-aware high-frequency energy | attention-based detector · RAGTruth and HalluRAG | Avg AUROC: 0.7627 -> 0.8052 (LLaMA-7B, entropy vs Fourier-high) | Improved |
| use Fourier-high instead of Lookback-Lens on summarization | hallucination detector · RAGTruth-Summ (LLaMA-13B) | AUROC: 0.6702 -> 0.7362 | Improved |
| use Fourier-high instead of Lookback-Lens on Mistral-7B QA | hallucination detector · RAGTruth-QA (Mistral-7B) | AUROC: 0.9148 -> 0.9099 | No clear effect |
| use Fourier operator instead of Wavelet or Laplacian (ablation) | frequency-aware features · RAGTruth + HalluRAG | Avg AUROC: 0.8052 / 0.7953 / 0.7859 (LLaMA-7B) | Mixed |
| restrict the classifier to the top 100 heads (ablation) | Fourier-high · RAGTruth-Avg (LLaMA-7B) | AUROC: 0.8205 -> 0.7915 | Hurt |
| restrict the classifier to the top 10 heads (ablation) | Fourier-high · RAGTruth-Avg (LLaMA-7B) | AUROC: 0.8205 -> 0.6995 | Hurt |
| remove context-attention features (ablation) | frequency-aware detector · RAGTruth and HalluRAG | AUROC | Hurt |
| aggregate features into 8-token spans instead of single tokens | Fourier-high · RAGTruth-Avg (LLaMA-7B) | AUROC: 0.8052 -> 0.8412 vs token-level avg (different task sets) | Improved |
| use Fourier-high instead of Lookback-Lens in cross-task transfer | hallucination detector · cross-domain (train on one task, test on another) | AUROC | Improved |
Closest papers by meaning
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.93)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.92)
- Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations (CVPR 2026 · similarity 0.92)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.91)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
