Research map / Large language models / Language, safety and interpretability
INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes INSIDE, which detects LLM hallucinations using an EigenScore (log-determinant of the covariance of multi-sample hidden embeddings) plus test-time feature clipping of extreme penultimate-layer activations.
Problem
Detecting knowledge hallucinations in LLM QA responses; existing token-logit uncertainty and language-level self-consistency methods lose semantic information during decoding.
Approach
Measure self-consistency of K sampled responses directly in the LLM's internal sentence-embedding space via the LogDet (eigenvalues) of their covariance matrix, which is linked to differential entropy, and clip extreme activations at test time to reduce overconfident hallucinations.
Key results
- EigenScore improves AUROC over Lexical Similarity by up to 5.6 points on CoQA and 8.9 points on SQuAD (LLaMA-7B/13B, OPT-6.7B), with best PCC in most settings.
- Feature clipping improves the base methods, with the largest AUROC gain of 1.8 points.
- On TruthfulQA with 50 prompts, EigenScore reaches 81.3 accuracy, versus 83.3 for ITI which is trained on TruthfulQA validation data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EigenScore instead of Lexical Similarity | LLaMA-7B hallucination detection · CoQA | AUROC (sentence similarity): 74.8 -> 80.4 | Helped |
| Use EigenScore instead of Lexical Similarity | LLaMA-7B hallucination detection · SQuAD | AUROC (sentence similarity): 74.9 -> 81.5 | Helped |
| Use EigenScore instead of Lexical Similarity | OPT-6.7B hallucination detection · SQuAD | AUROC (sentence similarity): 72.8 -> 81.7 | Helped |
| Use EigenScore instead of Perplexity | LLaMA-7B hallucination detection · TriviaQA | AUROC (sentence similarity): 83.6 -> 82.7 | Hurt |
| Add test-time feature clipping (ablation) | EigenScore with LLaMA-7B · CoQA | AUROC (sentence similarity): 79.3 -> 80.4 | Helped |
| Add test-time feature clipping (ablation) | EigenScore with OPT-6.7B · NQ | AUROC (sentence similarity): 77.1 -> 77.9 | Helped |
| Determine clipping thresholds with a memory bank rather than the current input sample (ablation) | EigenScore with LLaMA-7B · CoQA | AUROC (sentence similarity): 78.1 -> 80.4 | Helped |
| Determine clipping thresholds with a memory bank rather than a precomputed batch (ablation) | EigenScore with LLaMA-7B · NQ | AUROC (sentence similarity): 75.3 -> 76.5 | Helped |
| Increase the number of generations from below 15 toward 15 (ablation) | EigenScore, LN-Entropy, and Lexical Similarity · NQ | AUROC (sentence similarity) | Helped |
| Increase the number of generations beyond 15 (ablation) | EigenScore · NQ | AUROC (sentence similarity) | No clear effect |
| Use the last-token embedding from a middle layer instead of averaged token embeddings (ablation) | EigenScore with LLaMA-7B · CoQA | AUROC (sentence similarity) | Helped |
| Use sentence embeddings from shallow or final layers instead of layers near the middle (ablation) | EigenScore · CoQA | AUROC (sentence similarity) | Hurt |
Surprising findings
- Perplexity is best on TriviaQA but performs poorly on CoQA and SQuAD, attributed to very short TriviaQA answers.
- Consistency-based methods degrade sharply when sampling temperature exceeds 1.
Limitations
- Requires access to internal hidden states, so it cannot be applied to black-box models.
- Sampling-based approach requires K generations, adding inference cost.
- Only detects hallucinations; does not mitigate them.
Experimental setup
Datasets: CoQA SQuAD v2.0 Natural Questions TriviaQA TruthfulQA
Benchmarks: CoQA SQuAD NQ TriviaQA TruthfulQA
Metrics: AUROC PCC accuracy G-Mean
Baselines: Perplexity LN-Entropy Lexical Similarity Energy Semantic Entropy SentSAR SelfCheckGPT ITI
Builds on: Lexical Similarity Length-normalized Entropy SelfCheckGPT ReAct-style OOD feature rectification
Closest papers by meaning
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.92)
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.91)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.91)
- Steer LLM Latents for Hallucination Detection (ICML 2025 · similarity 0.90)
- ANAH-v2: Scaling Analytical Hallucination Annotation of Large Language Models (NeurIPS 2024 · similarity 0.90)
- Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.90)
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.