Hugging Face Daily Papers · September 1, 2026 · 4 min read

PaperGym: Rubric-Centered Evolution for Research-Plan Generation

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

We propose PaperGym, a low-leakage benchmark that turns arXiv papers into training environments by separating each paper's research question from its answer. Rubrics score proposals on methodological innovation and experimental design, and the same rubrics drive training via rubric-conditioned OPSD followed by rubric-rewarded GRPO.</p>\n","updatedAt":"2026-09-01T03:05:36.501Z","author":{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","fullname":"Zhengxi Lu","name":"LZXzju","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":13,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.897983193397522},"editors":["LZXzju"],"editorAvatarUrls":["/avatars/dfd802a24bd63e509728159ebb1769f6.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.31119","authors":[{"_id":"6a963e7fcd6ebc484732ec68","name":"Yuhan Wang","hidden":false},{"_id":"6a963e7fcd6ebc484732ec69","name":"Zhengxi Lu","hidden":false},{"_id":"6a963e7fcd6ebc484732ec6a","name":"Yuchen Yan","hidden":false},{"_id":"6a963e7fcd6ebc484732ec6b","name":"Kaitao Song","hidden":false},{"_id":"6a963e7fcd6ebc484732ec6c","name":"Wenqi Zhang","hidden":false},{"_id":"6a963e7fcd6ebc484732ec6d","name":"Weiming Lu","hidden":false},{"_id":"6a963e7fcd6ebc484732ec6e","name":"Jun Xiao","hidden":false},{"_id":"6a963e7fcd6ebc484732ec6f","name":"Yueting Zhuang","hidden":false},{"_id":"6a963e7fcd6ebc484732ec70","name":"Yongliang Shen","hidden":false}],"publishedAt":"2026-08-31T00:00:00.000Z","submittedOnDailyAt":"2026-09-01T00:00:00.000Z","title":"PaperGym: Rubric-Centered Evolution for Research-Plan Generation","submittedOnDailyBy":{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user","name":"LZXzju"},"summary":"Research planning is the decisive capability of AI scientists. Yet a research plan admits no verifiable answer, so reinforcement learning lacks the environment it requires: tasks paired with a critic. Rubrics extracted from scientific papers can supply the critic. Existing pipelines, however, draw the question and the criteria from the same content, so the reward can be earned by paraphrase. The rubric is further compressed into a single scalar per rollout. We introduce PaperGym, a unified framework that turns each research paper into a complete training environment. PaperGym exploits the structure of a paper: the question is synthesized from the research goal and background, while the criteria are derived from the method and experiments. The criteria span methodological innovation and experimental design, and criterion leakage falls to 3.7%, versus 11.90% to 34.10% in existing datasets. Training uses the rubric twice: first as privileged context for OPSD's self-teacher, then as the reward for GRPO. Across Qwen3-1.7B/4B/8B, this schedule outperforms supervised fine-tuning, either stage alone, and the reverse ordering, improving five-benchmark averages by +5.6, +5.0, and +4.8 points. With the recipe held fixed, models trained on PaperGym-20k win 58.1% of three-way comparisons, against 28.2% for RubricHub Science. The trained Qwen3-8B reaches 73.48 on ResearchQA, above the far larger Kimi K2.6. We release the pipeline, the 20,000-instance corpus PaperGym-20k, and the benchmarks PaperGym-Innov and PaperGym-Design.","upvotes":28,"discussionId":"6a963e80cd6ebc484732ec71","projectPage":"https://zju-real.github.io/PaperGym/","githubRepo":"https://github.com/ZJU-REAL/PaperGym","githubRepoAddedBy":"user","ai_summary":"PaperGym converts scientific papers into training environments by separating research questions from evaluation rubrics, enabling reinforcement learning that improves research planning across multiple model sizes.","ai_keywords":["PaperGym","reinforcement learning","rubric","self-teacher","GRPO","OPSD","criterion leakage","ResearchQA","PaperGym-20k"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":2},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"676127cf11b19ea602bb202a","avatarUrl":"/avatars/dfd802a24bd63e509728159ebb1769f6.svg","isPro":false,"fullname":"Zhengxi Lu","user":"LZXzju","type":"user"},{"_id":"6a82482fc969d4d85f255ab8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a82482fc969d4d85f255ab8/PkxtV_1sKNc6nC6QjtH3C.jpeg","isPro":false,"fullname":"Victor Wu","user":"wuvictor","type":"user"},{"_id":"6a7c2fa36c6c31e1230861df","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a7c2fa36c6c31e1230861df/hHPpEC90Yjt86wknlCuBK.jpeg","isPro":false,"fullname":"田中陽菜","user":"leijx3278","type":"user"},{"_id":"6a7c2058aa6c6e40950a3426","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a7c2058aa6c6e40950a3426/04vF3trA2yiqjdVFJgGkj.jpeg","isPro":false,"fullname":"高橋葉月","user":"adinugrohoport","type":"user"},{"_id":"6a8753dadd041ce77a69faa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a8753dadd041ce77a69faa7/HpBXFgiVIf6fWhcQ2cNGq.jpeg","isPro":false,"fullname":"Adam Stewart","user":"adamstewart","type":"user"},{"_id":"6a7db7fb2beeed89a7d16783","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a7db7fb2beeed89a7d16783/GhgYqcJDZEDxj4iW3Q5RM.jpeg","isPro":false,"fullname":"Phan Ngọc An","user":"Searoberts","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":"6a1fd9d481eee8267ebfe1f6","avatarUrl":"/avatars/4703ee72cb11285299f7fbc7128524a9.svg","isPro":false,"fullname":"Zhuowen Han","user":"Zhuowen02","type":"user"},{"_id":"66a0bec6b096327027e00667","avatarUrl":"/avatars/4c6ac70959f5e9d8d8e624b430cf678b.svg","isPro":false,"fullname":"Halen","user":"Zethive","type":"user"},{"_id":"6a8762050c6b6d23afbf7e50","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a8762050c6b6d23afbf7e50/GaLj4uzXH84GrXDqU1tCd.jpeg","isPro":false,"fullname":"Gabriele Santoro","user":"gabrielesantoro","type":"user"},{"_id":"6039478ab3ecf716b1a5fd4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6039478ab3ecf716b1a5fd4d/_Thy4E7taiSYBLKxEKJbT.jpeg","isPro":true,"fullname":"taesiri","user":"taesiri","type":"user"},{"_id":"6a7de3bcf03e08cfd11c422e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a7de3bcf03e08cfd11c422e/kjXaU2UdTu5W8X78e0TJv.jpeg","isPro":false,"fullname":"Olivia Brown","user":"SAMSONTSENG","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.31119.md","query":{}}">

Papers

arxiv:2608.31119

PaperGym: Rubric-Centered Evolution for Research-Plan Generation

Published on Aug 31

· Submitted by

Zhengxi Lu on Sep 1

Upvote

Authors:

Abstract

PaperGym converts scientific papers into training environments by separating research questions from evaluation rubrics, enabling reinforcement learning that improves research planning across multiple model sizes.

Generated by thinkingmachines/Inkling-Small

Research planning is the decisive capability of AI scientists. Yet a research plan admits no verifiable answer, so reinforcement learning lacks the environment it requires: tasks paired with a critic. Rubrics extracted from scientific papers can supply the critic. Existing pipelines, however, draw the question and the criteria from the same content, so the reward can be earned by paraphrase. The rubric is further compressed into a single scalar per rollout. We introduce PaperGym, a unified framework that turns each research paper into a complete training environment. PaperGym exploits the structure of a paper: the question is synthesized from the research goal and background, while the criteria are derived from the method and experiments. The criteria span methodological innovation and experimental design, and criterion leakage falls to 3.7%, versus 11.90% to 34.10% in existing datasets. Training uses the rubric twice: first as privileged context for OPSD's self-teacher, then as the reward for GRPO. Across Qwen3-1.7B/4B/8B, this schedule outperforms supervised fine-tuning, either stage alone, and the reverse ordering, improving five-benchmark averages by +5.6, +5.0, and +4.8 points. With the recipe held fixed, models trained on PaperGym-20k win 58.1% of three-way comparisons, against 28.2% for RubricHub Science. The trained Qwen3-8B reaches 73.48 on ResearchQA, above the far larger Kimi K2.6. We release the pipeline, the 20,000-instance corpus PaperGym-20k, and the benchmarks PaperGym-Innov and PaperGym-Design.