Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models
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.
Community
👋 Hi Hugging Face community! We’re excited to share our new paper:
🧠Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models
Large reasoning models generate long chains of thought, but are all reasoning steps useful for detecting hallucinations?
We find that reasoning traces frequently contain two types of noise:
🔸 Irrelevant steps that contribute little to solving the problem
🔸 Repetitive steps whose information is already captured elsewhere
These noisy steps can obscure the signals needed to determine whether the final answer is correct.
To address this, we introduce ReDe, a lightweight framework that:
✅ Uses final-answer attention as automatic supervision
✅ Learns representations that separate informative and noisy steps
✅ Filters reasoning traces without human step-level annotations
✅ Can be combined with probing-, uncertainty-, and verbalization-based detectors
📈 Across TruthfulQA, MATH, CodeElo, and MultiHopQA, ReDe consistently improves hallucination detection on Qwen3 and DeepSeek-R1 models. On TruthfulQA, it improves AUROC by up to 18.69 percentage points, reaching 87.32 AUROC.
Our main takeaway is simple:
For reliable reasoning models, we should not only analyze the reasoning trace—we should denoise it first.
📄 Paper: https://arxiv.org/abs/2607.22098
We would love to hear the community’s thoughts: What other downstream applications could benefit from filtering noisy reasoning steps?
Get this paper in your agent:
hf papers read 2607.22098 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper