Research map / Large language models / Reasoning and chain of thought
Mechanistic Detection and Mitigation of Hallucination in Large Reasoning Models
Zhongxiang Sun, Qipeng Wang, Haoyu Wang, Xiao Zhang, Jun Xu
Read the paperCheck your idea against 137,217 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.
Studies reasoning hallucination in large reasoning models, where logically coherent but factually wrong reasoning traces yield faulty conclusions, and proposes mechanistic detection and mitigation methods for it.
Problem
Large reasoning models produce reasoning hallucinations: logically coherent but factually incorrect multi-step traces that lead to persuasive but wrong answers, which are harder to detect than traditional hallucinations because they are embedded in structured reasoning.
Approach
Uses a mechanistic (internal-representation) analysis of LRMs to detect reasoning hallucinations and to mitigate them; the abstract does not specify the exact method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RHD (reasoning-score-based detector) instead of best baseline | R1-7B · ReTruthQA (MATH) | AUC: 0.7727 -> 0.7978 | Improved |
| use RHD instead of strongest baseline | R1-7B · ReTruthQA (MultiHopQA) | MC1: 0.6623 -> 0.7660 | Improved |
| use RHD instead of baselines | R1-14B · ReTruthQA (MATH) | MC1: 0.3846 -> 0.3692 | No clear effect |
| add reasoning-score reward shaping (GRPO-R) over GRPO | DeepSeek-R1-1.5B · MATH500 | accuracy: 0.770 -> 0.788 | Improved |
| add reasoning-score reward shaping (GRPO-R) over GRPO | DeepSeek-R1-1.5B · GPQA diamond | accuracy: 0.359 -> 0.414 | Improved |
| add reasoning-score reward shaping (GRPO-R) over GRPO | Qwen2.5-1.5B-Instruct · AIME 2024 | accuracy: 0.033 -> 0.133 | Improved |
| add reasoning-score reward shaping (GRPO-R) over GRPO | Qwen2.5-1.5B-Instruct · GPQA diamond | accuracy: 0.247 -> 0.247 | No clear effect |
| plain GRPO versus base model | DeepSeek-R1-1.5B · MATH500 | accuracy: 0.772 -> 0.770 | No clear effect |
| GRPO-R vs GRPO on GPQA extended | DeepSeek-R1-1.5B · GPQA extended | accuracy: 0.359 -> 0.357 | Hurt |
Closest papers by meaning
- Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.93)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.93)
- Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.92)
- MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM (NeurIPS 2025 · similarity 0.92)
- More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models (NeurIPS 2025 · similarity 0.92)
- Learning to Reason for Hallucination Span Detection (ICLR 2026 · 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.
