Research map / Large language models / Language, safety and interpretability
Automatic Layer Selection for Hallucination Detection
Xinpeng Wang, William Cao, Andrew Gordon Wilson, Zhe Zeng
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 first-sentence truncation (FST) token position (ablation) | FEPoID probing · QA average, Llama-3.1-8B-Instruct | AUROC: 0.7253 -> 0.8752 | Improved |
| add first-sentence truncation (FST) token position (ablation) | FEPoID probing · QA average, Mistral-7B-Instruct-v0.3 | AUROC: 0.8531 -> 0.8687 | Improved |
| add first-sentence truncation (FST) token position (ablation) | FEPoID probing · QA average, Llama-3.1-8B (base) | AUROC: 0.7895 -> 0.8581 | Improved |
| add first-sentence truncation (FST) token position (ablation) | FEPoID probing · QA average, Llama-3.2-3B | AUROC: 0.7542 -> 0.8339 | Improved |
| add first-sentence truncation (FST) token position (ablation) | FEPoID probing · QA average, Llama-3.2-1B | AUROC: 0.7061 -> 0.7860 | Improved |
| add FST (ablation) | Lexical Similarity · QA average, Llama-3.1-8B-Instruct | AUROC: 0.6196 -> 0.8187 | Improved |
| add FST (ablation) | Predictive Entropy · QA average, Llama-3.1-8B-Instruct | AUROC: 0.5820 -> 0.7638 | Improved |
| add FST (ablation) | Fisher Separation of representations · CoQA, Llama-3.1-8B-Instruct | Fisher Separation: 0.0003 -> 0.0044 | Improved |
| add FST (ablation) | Silhouette Score of representations · PsiLoQA, Llama-3.1-8B-Instruct | Silhouette Score: 0.0118 -> 0.2605 | Improved |
| add first-effective-peak filtering instead of max ID layer (ablation) | ID-based layer selection · QA average, Llama-3.1-8B-Instruct | AUROC: 0.6843 -> 0.7253 | Improved |
| add first-effective-peak filtering instead of max ID layer (ablation) | ID-based layer selection · QA average, Mistral-7B-Instruct-v0.3 | AUROC: 0.8071 -> 0.8531 | Improved |
| add first-effective-peak filtering instead of max ID layer (ablation) | ID-based layer selection · Summarization average, Llama-3.1-8B-Instruct | AUROC: 0.5997 -> 0.6080 | 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 0.93)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.92)
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.92)
- Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (NeurIPS 2026 · similarity 0.92)
- HalLoc: Token-level Localization of Hallucinations for Vision Language Models (CVPR 2025 · similarity 0.92)
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.
