Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models

2026-07-24

AI Summary

This work addresses the challenge that irrelevant and repetitive steps in long reasoning trajectories hinder hallucination detection. To tackle this issue, the authors propose REDE, a novel framework that leverages the attention distribution over the final answer to automatically generate denoising supervision signals—eliminating the need for manual annotations while optimizing step-level representation spaces. REDE achieves plug-and-play reasoning trajectory purification through attention-guided representation learning, embedding space reconstruction, and noisy-step filtering. Extensive experiments across multiple reasoning benchmarks demonstrate that REDE substantially outperforms existing methods and consistently enhances hallucination detection performance.

Abstract

Large reasoning models (LRMs) generate long reasoning traces before producing final answers. While these traces may contain useful signals for hallucination detection, harnessing them is non-trivial because long trajectories often include noisy steps that obscure the cues relevant to truthfulness assessment. In this paper, we identify two prevalent forms of reasoning noises, i.e., irrelevant steps and repetitive steps, and show that both substantially degrade hallucination detection performance. Existing confidence-based scores and naive embedding-based filtering fail to reliably separate noisy from informative steps. To address this challenge, we propose REDE, a novel learning framework for denoising reasoning traces for hallucination detection. Specifically, REDE leverages final-answer attention as an automatic supervision signal to shape the step-level representation space, yielding refined embeddings in which noisy steps can be reliably identified and filtered. REDE can be readily plugged into diverse hallucination detectors by operating on the filtered reasoning trajectory after removing noisy steps. Extensive experiments on multiple reasoning benchmarks show that REDE consistently improves detection performance over competitive baselines.

Problem

• reasoning traces
• hallucination detection
• reasoning noise
• large reasoning models

Innovation

• reasoning denoising
• hallucination detection
• large reasoning models
• attention-based supervision
• step-level representation