Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models
TL;DR AI
2 min readKey summary
Researchers introduced ReDe, a lightweight framework for denoising reasoning traces and separating useful steps from noisy ones.
It uses final-answer attention as supervision, avoiding step-level annotations while filtering repetitive or irrelevant reasoning.
Across TruthfulQA, MATH, CodeElo, and MultiHopQA with Qwen3 and DeepSeek-R1, it improved hallucination detection.
On TruthfulQA, ReDe delivered an AUROC gain of up to 18.69 points, suggesting more reliable model judgments.