AI Paper Atlas

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

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency

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 changedWhereMeasuredEffect
use RHD (reasoning-score-based detector) instead of best baselineR1-7B · ReTruthQA (MATH)AUC: 0.7727 -> 0.7978Improved
use RHD instead of strongest baselineR1-7B · ReTruthQA (MultiHopQA)MC1: 0.6623 -> 0.7660Improved
use RHD instead of baselinesR1-14B · ReTruthQA (MATH)MC1: 0.3846 -> 0.3692No clear effect
add reasoning-score reward shaping (GRPO-R) over GRPODeepSeek-R1-1.5B · MATH500accuracy: 0.770 -> 0.788Improved
add reasoning-score reward shaping (GRPO-R) over GRPODeepSeek-R1-1.5B · GPQA diamondaccuracy: 0.359 -> 0.414Improved
add reasoning-score reward shaping (GRPO-R) over GRPOQwen2.5-1.5B-Instruct · AIME 2024accuracy: 0.033 -> 0.133Improved
add reasoning-score reward shaping (GRPO-R) over GRPOQwen2.5-1.5B-Instruct · GPQA diamondaccuracy: 0.247 -> 0.247No clear effect
plain GRPO versus base modelDeepSeek-R1-1.5B · MATH500accuracy: 0.772 -> 0.770No clear effect
GRPO-R vs GRPO on GPQA extendedDeepSeek-R1-1.5B · GPQA extendedaccuracy: 0.359 -> 0.357Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy