Large reasoning models (LRMs) generate long chains of thought before arriving at final answers. While these reasoning traces contain valuable signals for detecting hallucinations—incorrect or fabricated outputs—they are often polluted by noisy steps that obscure truthfulness cues. In a new paper, researchers from Hugging Face and collaborators introduce REDE (Reasoning Denoiser), a learning framework designed to filter out such noise automatically.
The team identifies two primary types of reasoning noise: irrelevant steps that contribute little to solving the problem, and repetitive steps whose information is already captured elsewhere. Both types degrade the performance of existing hallucination detectors, which rely on confidence scores or naive embedding-based filtering that fail to reliably separate noisy from informative steps.
REDE leverages final-answer attention as an automatic supervision signal to shape step-level representations. This allows the model to learn embeddings where noisy steps can be identified and removed without requiring human step-level annotations. The filtered reasoning trajectory can then be fed into any existing hallucination detector—whether probing-based, uncertainty-based, or verbalization-based—to improve its accuracy.
Experiments on benchmarks including TruthfulQA, MATH, CodeElo, and MultiHopQA show consistent improvements across Qwen3 and DeepSeek-R1 models. On TruthfulQA, REDE boosts AUROC by up to 18.69 percentage points, reaching 87.32 AUROC. The paper is available on arXiv and has been shared with the Hugging Face community for feedback.