Paper link: <a href=\"https://arxiv.org/abs/2606.18910\" rel=\"nofollow\">https://arxiv.org/abs/2606.18910</a></p>\n","updatedAt":"2026-06-18T23:18:11.749Z","author":{"_id":"6719ddbdf639bea02ae5b1d6","avatarUrl":"/avatars/c18d2d7920fecf37d46b4f42cb6a444a.svg","fullname":"Yuanxin","name":"Yuanxin-Liu","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":0,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7800112366676331},"editors":["Yuanxin-Liu"],"editorAvatarUrls":["/avatars/c18d2d7920fecf37d46b4f42cb6a444a.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.18910","authors":[{"_id":"6a33917159127a45e2c1c6c9","user":{"_id":"6719ddbdf639bea02ae5b1d6","avatarUrl":"/avatars/c18d2d7920fecf37d46b4f42cb6a444a.svg","isPro":true,"fullname":"Yuanxin","user":"Yuanxin-Liu","type":"user","name":"Yuanxin-Liu"},"name":"Yuanxin Liu","status":"claimed_verified","statusLastChangedAt":"2026-06-18T11:26:28.631Z","hidden":false},{"_id":"6a33917159127a45e2c1c6ca","name":"Ruida Zhou","hidden":false},{"_id":"6a33917159127a45e2c1c6cb","name":"Xinyan Zhao","hidden":false},{"_id":"6a33917159127a45e2c1c6cc","name":"Amr Sharaf","hidden":false},{"_id":"6a33917159127a45e2c1c6cd","name":"Hongzhou Lin","hidden":false},{"_id":"6a33917159127a45e2c1c6ce","name":"Arijit Biswas","hidden":false},{"_id":"6a33917159127a45e2c1c6cf","name":"Mohammad Ghavamzadeh","hidden":false},{"_id":"6a33917159127a45e2c1c6d0","name":"Zhaoran Wang","hidden":false},{"_id":"6a33917159127a45e2c1c6d1","name":"Mingyi Hong","hidden":false}],"publishedAt":"2026-06-17T00:00:00.000Z","submittedOnDailyAt":"2026-06-18T00:00:00.000Z","title":"REVES: REvision and VErification--Augmented Training for Test-Time Scaling","submittedOnDailyBy":{"_id":"6719ddbdf639bea02ae5b1d6","avatarUrl":"/avatars/c18d2d7920fecf37d46b4f42cb6a444a.svg","isPro":true,"fullname":"Yuanxin","user":"Yuanxin-Liu","type":"user","name":"Yuanxin-Liu"},"summary":"Test-time scaling via sequential revision has emerged as a powerful paradigm for enhancing Large Language Model (LLM) reasoning. However, standard post-training methods primarily optimize single-shot objectives, creating a fundamental misalignment with multi-step inference dynamics. While recent work treats this as multi-turn reinforcement learning (RL), conventional approaches optimize over the multi-step trajectories directly, failing to further exploit the high-quality mistakes in intermediate steps that model can learn from correcting them. We propose a two-stage iterative framework that alternates between online data/prompt augmentation and policy optimization. By converting the intermediate steps (``near-miss'' answers) in the successful recovery trajectories into decoupled revision and verification prompts, our approach concentrates training on both effective answer transformation and error identification. This approach enables efficient off-policy data generation and reduces the computational overhead of long-horizon sampling compared to standard multi-turn RL. On LiveCodeBench, using publicly available test cases as feedback, we observe gains of +6.5 points over the RL baseline and +4.0 points over standard multi-turn training. Beyond coding, our approach matches the previously reported SOTA result on circle packing while using the smallest base model (4B) and far fewer rollouts than the much larger evolutionary search systems. Math results under ground-truth verification further confirm improved correction ability. It also generalizes to out-of-distribution constraint-satisfaction puzzles such as n\\_queens and mini\\_sudoku, where correctness is defined entirely by problem constraints. Code is available at https://github.com/yxliu02/REVES.git.","upvotes":2,"discussionId":"6a33917159127a45e2c1c6d2","githubRepo":"https://github.com/yxliu02/REVES","githubRepoAddedBy":"user","ai_summary":"A two-stage iterative framework alternates between data augmentation and policy optimization to improve LLM reasoning by leveraging intermediate correction steps, achieving superior performance on coding benchmarks and constraint satisfaction problems.","ai_keywords":["Large Language Model","multi-step inference dynamics","reinforcement learning","off-policy data generation","test-time scaling","sequential revision","prompt augmentation","policy optimization","intermediate steps","error identification","answer transformation"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":0},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6719ddbdf639bea02ae5b1d6","avatarUrl":"/avatars/c18d2d7920fecf37d46b4f42cb6a444a.svg","isPro":true,"fullname":"Yuanxin","user":"Yuanxin-Liu","type":"user"},{"_id":"69195d6a4311763d9d8b1cb5","avatarUrl":"/avatars/32055830f1a5c72393fcf6452e894358.svg","isPro":true,"fullname":"Liam","user":"lyx02klmy","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.18910.md","query":{}}">
REVES: REvision and VErification--Augmented Training for Test-Time Scaling
Abstract
A two-stage iterative framework alternates between data augmentation and policy optimization to improve LLM reasoning by leveraging intermediate correction steps, achieving superior performance on coding benchmarks and constraint satisfaction problems.
Test-time scaling via sequential revision has emerged as a powerful paradigm for enhancing Large Language Model (LLM) reasoning. However, standard post-training methods primarily optimize single-shot objectives, creating a fundamental misalignment with multi-step inference dynamics. While recent work treats this as multi-turn reinforcement learning (RL), conventional approaches optimize over the multi-step trajectories directly, failing to further exploit the high-quality mistakes in intermediate steps that model can learn from correcting them. We propose a two-stage iterative framework that alternates between online data/prompt augmentation and policy optimization. By converting the intermediate steps (``near-miss'' answers) in the successful recovery trajectories into decoupled revision and verification prompts, our approach concentrates training on both effective answer transformation and error identification. This approach enables efficient off-policy data generation and reduces the computational overhead of long-horizon sampling compared to standard multi-turn RL. On LiveCodeBench, using publicly available test cases as feedback, we observe gains of +6.5 points over the RL baseline and +4.0 points over standard multi-turn training. Beyond coding, our approach matches the previously reported SOTA result on circle packing while using the smallest base model (4B) and far fewer rollouts than the much larger evolutionary search systems. Math results under ground-truth verification further confirm improved correction ability. It also generalizes to out-of-distribution constraint-satisfaction puzzles such as n\_queens and mini\_sudoku, where correctness is defined entirely by problem constraints. Code is available at https://github.com/yxliu02/REVES.git.
Community
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2606.18910 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.18910 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.18910 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.