Research map / Large language models / Retrieval and knowledge
Attributive Reasoning for Hallucination Diagnosis of Large Language Models
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 an attribution framework that links LLM hallucination categories to internal signals, builds the RelQA-Cate benchmark (eight hallucination categories), and introduces Differential Penalty Decoding (DPD) that reweights candidate answers by internal-state penalties.
Problem
Hallucinations in LLMs are typically treated without distinguishing their categories or examining the internal states that cause them; existing mitigation methods (retrieval, prompting, data, RL) do not diagnose categories or internal causes.
Approach
Analyze hidden-layer activations, attention outputs, and gradient-based salient words to attribute hallucinations to categories, then convert these differences from correct answers into penalty terms that down-weight hallucinated candidate answers during decoding.
Key results
- DPD improves LLaMA-7B on RelQA-Cate F1 from 52.21 (DoLa) to 57.54 and on TruthfulQA MC2 from 40.24 to 44.26.
- On LLaMA2-7B, DPD reaches 74.28 F1 on RelQA-Cate and 49.63 MC1 on TruthfulQA, outperforming RHO and FLEEK.
- The paper reports improvements up to 28.25% in answer reliability, with largest relative gain of 30.62% for Mistral 8x7B on TruthfulQA MC1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DPD decoding instead of the previous decoding strategy | LLaMA-7B · TruthfulQA | MC1: 31.95 -> 34.27 | Helped |
| Apply DPD decoding instead of the previous decoding strategy | LLaMA2-7B · RelQA-Cate | F1: 42.77 -> 48.55 | Helped |
| Apply DPD decoding instead of the previous decoding strategy | Baichuan2-7B · TruthfulQA | MC3: 39.67 -> 44.30 | Helped |
| Apply DPD decoding instead of the previous decoding strategy | Mistral-7B · TruthfulQA | MC2: 74.73 -> 78.35 | Helped |
| Increase candidate-answer diversity beyond 0.8 (ablation) | DPD · RelQA-Cate | F1 | Helped |
| Add the hidden-layer output penalty Pl (ablation) | DPD · RelQA-Cate | F1 | Helped |
| Add the PCA-distance penalty Pdl (ablation) | DPD · RelQA-Cate | F1 | Helped |
| Add the attention-output penalty Pa (ablation) | DPD · RelQA-Cate | F1 | Helped |
| Add the attention-correlation penalty Pca (ablation) | DPD · RelQA-Cate | F1 | No clear effect |
| Add the high-contribution-word penalty Pwa (ablation) | DPD · RelQA-Cate | F1 | Helped |
Surprising findings
- Removing the context-independent penalty Pca produced almost no performance decrease, while removing the high-contribution word penalty Pwa caused the largest drop.
- Improvements were smaller for Baichuan2-7B and Mistral-7B than for the LLaMA series, which the authors attribute to stronger baseline QA performance.
Limitations
- Penalty coefficients and the global alpha are tuned per dataset, and penalties are tailored to each dataset, limiting generality.
- Hallucination category labels were produced by ChatGPT and human validation, and categories may overlap.
- Ground-truth answers are required to compute penalty baselines; the paper only sketches alternatives (GPT-4 or a reward model) without evaluating them.
Experimental setup
Datasets: RelQA TruthfulQA
Benchmarks: TruthfulQA RelQA-Cate
Metrics: MC1 MC2 MC3 F1 CGS (ChatGPT score)
Baselines: Alpaca ITI CD DoLa SH2 ICD RARR CoVe CoNLI RHO FLEEK
Builds on: DoLa Distinct-2 ITI gradient saliency maps
Closest papers by meaning
- HaDeMiF: Hallucination Detection and Mitigation in Large Language Models (ICLR 2025 · similarity 0.93)
- LLM-Check: Investigating Detection of Hallucinations in Large Language Models (NeurIPS 2024 · similarity 0.92)
- Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations (NeurIPS 2025 · similarity 0.92)
- Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.92)
- In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.92)
- Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 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.