For reliable reasoning models, we should not only analyze the reasoning trace—we should denoise it first.\n\n📄 Paper: https://arxiv.org/abs/2607.22098\n\nWe would love to hear the community’s thoughts: **What other downstream applications could benefit from filtering noisy reasoning steps?**\n","html":"<p>👋 Hi Hugging Face community! We’re excited to share our new paper:</p>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"🧠-reasoning-denoiser-denoising-reasoning-traces-for-hallucination-detection-in-large-reasoning-models\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#🧠-reasoning-denoiser-denoising-reasoning-traces-for-hallucination-detection-in-large-reasoning-models\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t🧠 Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models\n\t</span>\n</h2>\n<p>Large reasoning models generate long chains of thought, but are all reasoning steps useful for detecting hallucinations?</p>\n<p>We find that reasoning traces frequently contain two types of noise:</p>\n<p>🔸 <strong>Irrelevant steps</strong> that contribute little to solving the problem<br>🔸 <strong>Repetitive steps</strong> whose information is already captured elsewhere</p>\n<p>These noisy steps can obscure the signals needed to determine whether the final answer is correct.</p>\n<p>To address this, we introduce <strong>ReDe</strong>, a lightweight framework that:</p>\n<p>✅ Uses final-answer attention as automatic supervision<br>✅ Learns representations that separate informative and noisy steps<br>✅ Filters reasoning traces without human step-level annotations<br>✅ Can be combined with probing-, uncertainty-, and verbalization-based detectors</p>\n<p>📈 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 <strong>18.69 percentage points</strong>, reaching <strong>87.32 AUROC</strong>.</p>\n<p>Our main takeaway is simple:</p>\n<blockquote>\n<p>For reliable reasoning models, we should not only analyze the reasoning trace—we should denoise it first.</p>\n</blockquote>\n<p>📄 Paper: <a href=\"https://arxiv.org/abs/2607.22098\" rel=\"nofollow\">https://arxiv.org/abs/2607.22098</a></p>\n<p>We would love to hear the community’s thoughts: <strong>What other downstream applications could benefit from filtering noisy reasoning steps?</strong></p>\n","updatedAt":"2026-07-28T07:50:26.197Z","author":{"_id":"63374d7a0267ebcf0266c83d","avatarUrl":"/avatars/dab490fc1950f2778f5a6e9bf9893aaa.svg","fullname":"Sean Du","name":"xfdu1","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.5936712026596069},"editors":["xfdu1"],"editorAvatarUrls":["/avatars/dab490fc1950f2778f5a6e9bf9893aaa.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.22098","authors":[{"_id":"6a685cef81e96813d6f78963","name":"Junlin Fang","hidden":false},{"_id":"6a685cef81e96813d6f78964","name":"Do Nguyen-Thanh","hidden":false},{"_id":"6a685cef81e96813d6f78965","name":"Xiaogang Xu","hidden":false},{"_id":"6a685cef81e96813d6f78966","name":"Zhen Fang","hidden":false},{"_id":"6a685cef81e96813d6f78967","name":"Sean Du","hidden":false}],"publishedAt":"2026-07-24T00:00:00.000Z","submittedOnDailyAt":"2026-07-28T00:00:00.000Z","title":"Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models","submittedOnDailyBy":{"_id":"63374d7a0267ebcf0266c83d","avatarUrl":"/avatars/dab490fc1950f2778f5a6e9bf9893aaa.svg","isPro":false,"fullname":"Sean Du","user":"xfdu1","type":"user","name":"xfdu1"},"summary":"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.","upvotes":5,"discussionId":"6a685cef81e96813d6f78968","organization":{"_id":"6508b28cf36bb51c50faad98","name":"NanyangTechnologicalUniversity","fullname":"Nanyang Technological University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/630ca0817dacb93b33506ce7/ZPD1fvei0bcIGeDXxeSkn.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"672b0c078573e4606eb5f8fe","avatarUrl":"/avatars/f4c2a86e995e32e22a6bf7288fe91068.svg","isPro":false,"fullname":"Junlin Fang","user":"Sol45","type":"user"},{"_id":"63374d7a0267ebcf0266c83d","avatarUrl":"/avatars/dab490fc1950f2778f5a6e9bf9893aaa.svg","isPro":false,"fullname":"Sean Du","user":"xfdu1","type":"user"},{"_id":"68416ddcafc844c48a921225","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/CkamXjlcuth0abSWY2EGu.png","isPro":false,"fullname":"Aditya Goyal","user":"AdityaGtheOg","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"661ab1f1fa3b144a381fa454","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png","isPro":false,"fullname":"Urro","user":"urroxyz","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6508b28cf36bb51c50faad98","name":"NanyangTechnologicalUniversity","fullname":"Nanyang Technological University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/630ca0817dacb93b33506ce7/ZPD1fvei0bcIGeDXxeSkn.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.22098.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2607.22098 in a model README.md to link it from this page.

Cite arxiv.org/abs/2607.22098 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2607.22098 in a Space README.md to link it from this page.