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

The paper argues that noisy reasoning steps make hallucination detection worse, then proposes a learned filter for those traces.

The authors identify irrelevant and repetitive steps as common noise in long reasoning traces from large reasoning models. Their method, REDE, uses final-answer attention as supervision to learn step-level embeddings that separate useful steps from noise. The filtered trace can then be fed into existing hallucination detectors. Experiments across multiple reasoning benchmarks are reported to improve detection performance over competitive baselines. HF Daily Papers' note