Research map / Large language models / Language, safety and interpretability
Efficient Hallucination Detection for LLMs Using Uncertainty-Aware Attention Heads
Artem Vazhentsev, Lyudmila Rvanova, Gleb Kuzmin, Ekaterina Fadeeva, Ivan Lazichny, Alexander Panchenko, Maxim Panov, Mrinmaya Sachan, Preslav Nakov, Timothy Baldwin, Artem Shelmanov
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 |
|---|---|---|---|
| add recurrence over previous-token confidence (ablation) | RAUQ · 12 datasets | PRR: .388 -> .413 | Improved |
| add attention weights to recurrent formula (ablation) | RAUQ · 12 datasets | PRR: .330 -> .413 | Improved |
| use mean of logs token aggregation instead of sum of logs (ablation) | RAUQ · 12 datasets | PRR: .198 -> .413 | Improved |
| use max instead of mean layer aggregation (ablation) | RAUQ · 12 datasets | PRR: .404 -> .413 | No clear effect |
| exclude prompt tokens from Attention Score (ablation) | Attention Score · 12 datasets | PRR: .069 -> .101 | Improved |
| add uncertainty-aware head selection to Attention Score (ablation) | Attention Score (Gen. Tokens) · 12 datasets | PRR: .101 -> .144 | Improved |
| add full RAUQ recurrence to Attention Score with heads (ablation) | Attention Score (Gen. Tokens, Selected Head) · 12 datasets | PRR: .144 -> .413 | Improved |
| use entropy instead of probability confidence (ablation) | RAUQ · Instruct model, 11 datasets | PRR: .367 -> .400 | Improved |
| use entropy instead of probability confidence (ablation) | RAUQ · Llama-3.1 8B base | PRR: .388 -> .391 | No clear effect |
| add change-point attention normalization (ablation) | RAUQ · 12 datasets | PRR: .413 -> .366 | Hurt |
| use RAUQ instead of Semantic Entropy | Llama-3.1 8B · runtime per instance | seconds: 6.40 -> 1.17 | Improved |
| replace attention weights with Layer Integrated Gradients (ablation) | RAUQ · 12 datasets | PRR: 0.4% drop | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Semantic Density — Semantic Density: Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.93)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.92)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.91)
- HaloProbe: Bayesian Detection and Mitigation of Object Hallucinations in Vision-Language Models (ICML 2026 · similarity 0.91)
- PAS: Prelim Attention Score for Detecting Object Hallucinations in Large Vision-Language Models (CVPR 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.
