<a href=\"https://cdn-uploads.huggingface.co/production/uploads/6520621836008ecc88699622/SkxGJh0u64gB1zZDE-VeA.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/6520621836008ecc88699622/SkxGJh0u64gB1zZDE-VeA.png\" alt=\"motiv\"></a></p>\n<p>Why Do Reasoning Models Lose Coverage?<br>We revisit this open question with a data-centric lens , and show how “forks in the road” situations in the post-training data can shape—and shrink—model coverage</p>\n","updatedAt":"2026-05-20T16:51:32.315Z","author":{"_id":"6520621836008ecc88699622","avatarUrl":"/avatars/b08c00af00f1736a4f4938443e575b0e.svg","fullname":"Parshin Shojaee","name":"parshinsh","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.755286693572998},"editors":["parshinsh"],"editorAvatarUrls":["/avatars/b08c00af00f1736a4f4938443e575b0e.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.17026","authors":[{"_id":"6a0de46dd1ef9ecdf71c0e1a","name":"Ngoc-Hieu Nguyen","hidden":false},{"_id":"6a0de46dd1ef9ecdf71c0e1b","user":{"_id":"6520621836008ecc88699622","avatarUrl":"/avatars/b08c00af00f1736a4f4938443e575b0e.svg","isPro":false,"fullname":"Parshin Shojaee","user":"parshinsh","type":"user","name":"parshinsh"},"name":"Parshin Shojaee","status":"claimed_verified","statusLastChangedAt":"2026-05-20T17:09:22.462Z","hidden":false},{"_id":"6a0de46dd1ef9ecdf71c0e1c","name":"Phuc Minh Nguyen","hidden":false},{"_id":"6a0de46dd1ef9ecdf71c0e1d","name":"Nan Zhang","hidden":false},{"_id":"6a0de46dd1ef9ecdf71c0e1e","name":"Chandan K Reddy","hidden":false},{"_id":"6a0de46dd1ef9ecdf71c0e1f","name":"Khoa D Doan","hidden":false},{"_id":"6a0de46dd1ef9ecdf71c0e20","name":"Rui Zhang","hidden":false}],"publishedAt":"2026-05-16T00:00:00.000Z","submittedOnDailyAt":"2026-05-20T00:00:00.000Z","title":"Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the Road","submittedOnDailyBy":{"_id":"6520621836008ecc88699622","avatarUrl":"/avatars/b08c00af00f1736a4f4938443e575b0e.svg","isPro":false,"fullname":"Parshin Shojaee","user":"parshinsh","type":"user","name":"parshinsh"},"summary":"Recent progress in large language models has led to the emergence of reasoning models, which have shown strong performance on complex tasks through specialized fine-tuning procedures. While these methods reliably improve pass@1 accuracy, prior works have observed that they show a coverage shrinkage behavior, where pass@k degrades relative to the base model. In this paper, we investigate the reasoning shrinkage arise under SFT-based post-training. We hypothesize that this behavior is driven by properties of the fine-tuning data, specifically related to decision points or \"forks in the road\" scenarios where model faces indecipherable patterns with multiple valid reasoning paths. To test this hypothesis, we design controlled case studies that simulate such decision-point settings, spanning indecipherable nodes in graph branching, and reasoning modes. By tracking post-training dynamics in these settings, we find that the shrinkage phenomenon is tightly correlated with the prevalence of decision-point scenarios in the training data. We also demonstrate that this shrinkage behavior can be partially mitigated through targeted data synthesis design of decision-points, and a more systematic diversity-encouraging decoding mechanism. Our findings identify data-centric factors as a key driver of shrinkage in reasoning models and highlight diversity-aware designs as an effective lever for controlling it.","upvotes":0,"discussionId":"6a0de46dd1ef9ecdf71c0e21","projectPage":"https://nnhieu.github.io/blog/2026/reasoning-forks/","githubRepo":"https://github.com/psunlpgroup/reasoning_forks","githubRepoAddedBy":"user","ai_summary":"Reasoning models exhibit coverage shrinkage during supervised fine-tuning due to decision-point scenarios in training data, which can be mitigated through targeted data synthesis and diversity-encouraging decoding mechanisms.","ai_keywords":["large language models","reasoning models","fine-tuning","pass@1 accuracy","pass@k","SFT-based post-training","decision points","forks in the road","graph branching","reasoning modes","data-centric factors","targeted data synthesis","diversity-encouraging decoding"],"githubStars":1},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["en"],"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.17026.md"}">
Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the Road
Abstract
Reasoning models exhibit coverage shrinkage during supervised fine-tuning due to decision-point scenarios in training data, which can be mitigated through targeted data synthesis and diversity-encouraging decoding mechanisms.
AI-generated summary
Recent progress in large language models has led to the emergence of reasoning models, which have shown strong performance on complex tasks through specialized fine-tuning procedures. While these methods reliably improve pass@1 accuracy, prior works have observed that they show a coverage shrinkage behavior, where pass@k degrades relative to the base model. In this paper, we investigate the reasoning shrinkage arise under SFT-based post-training. We hypothesize that this behavior is driven by properties of the fine-tuning data, specifically related to decision points or "forks in the road" scenarios where model faces indecipherable patterns with multiple valid reasoning paths. To test this hypothesis, we design controlled case studies that simulate such decision-point settings, spanning indecipherable nodes in graph branching, and reasoning modes. By tracking post-training dynamics in these settings, we find that the shrinkage phenomenon is tightly correlated with the prevalence of decision-point scenarios in the training data. We also demonstrate that this shrinkage behavior can be partially mitigated through targeted data synthesis design of decision-points, and a more systematic diversity-encouraging decoding mechanism. Our findings identify data-centric factors as a key driver of shrinkage in reasoning models and highlight diversity-aware designs as an effective lever for controlling it.
Community
Why Do Reasoning Models Lose Coverage?
We revisit this open question with a data-centric lens , and show how “forks in the road” situations in the post-training data can shape—and shrink—model coverage
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/2605.17026 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.17026 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.17026 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.