We present <strong>GoLongRL</strong>, a fully open-source, capability-oriented post-training recipe for long-context reinforcement learning with verifiable rewards (RLVR) 🚀</p>\n<p><strong>GoLongRL-30B-A3B</strong> achieves long-context performance comparable to DeepSeek-R1-0528 and Qwen3-235B-A22B-Thinking-2507—while using a significantly smaller activated parameter budget ⚡</p>\n<p>Under the same vanilla GRPO setup, our dataset alone outperforms the QwenLong-L1.5 dataset by <strong>+6.1pp at the 4B scale</strong> and <strong>+2.6pp at 30B</strong> 📈</p>\n<p>🔥 <strong>Actually open-source — not “weights-only”, not partial releases</strong>:<br>we release <strong>training data, full code, and models</strong>. No hidden pipelines, no inaccessible datasets, no black boxes.<br>If you want to reproduce, audit, or build on top of it—you <em>can</em>. End to end.</p>\n<p><strong>Resources:</strong></p>\n<ul>\n<li><p>📄 Paper: <a href=\"https://arxiv.org/pdf/2605.19577\" rel=\"nofollow\">https://arxiv.org/pdf/2605.19577</a></p>\n</li>\n<li><p>🧠 Github Project: <a href=\"https://github.com/xiaoxuanNLP/GoLongRL\" rel=\"nofollow\">https://github.com/xiaoxuanNLP/GoLongRL</a></p>\n</li>\n<li><p>📊 Data: <a href=\"https://huggingface.co/datasets/Kwai-Klear/GoLongRL\">https://huggingface.co/datasets/Kwai-Klear/GoLongRL</a></p>\n</li>\n<li><p>🤖 Models:</p>\n<ul>\n<li>GoLongRL-30B-A3B: <a href=\"https://huggingface.co/Kwai-Klear/GoLongRL-30B-A3B\">https://huggingface.co/Kwai-Klear/GoLongRL-30B-A3B</a></li>\n<li>GoLongRL-4B: <a href=\"https://huggingface.co/Kwai-Klear/GoLongRL-4B\">https://huggingface.co/Kwai-Klear/GoLongRL-4B</a></li>\n</ul>\n</li>\n</ul>\n","updatedAt":"2026-05-20T04:07:40.793Z","author":{"_id":"61c2cf8d1172fa7969904d99","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/61c2cf8d1172fa7969904d99/R10G5h3d9Q_YQ__Hc-H4k.jpeg","fullname":"suu","name":"Suu","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":5,"isUserFollowing":false}},"numEdits":5,"identifiedLanguage":{"language":"en","probability":0.7899413108825684},"editors":["Suu"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/61c2cf8d1172fa7969904d99/R10G5h3d9Q_YQ__Hc-H4k.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.19577","authors":[{"_id":"6a0d2af265eb30f20d962ca6","name":"Minxuan Lv","hidden":false},{"_id":"6a0d2af265eb30f20d962ca7","name":"Tiehua Mei","hidden":false},{"_id":"6a0d2af265eb30f20d962ca8","name":"Tanlong Du","hidden":false},{"_id":"6a0d2af265eb30f20d962ca9","name":"Junmin Chen","hidden":false},{"_id":"6a0d2af265eb30f20d962caa","name":"Zhenpeng Su","hidden":false},{"_id":"6a0d2af265eb30f20d962cab","name":"Ziyang Chen","hidden":false},{"_id":"6a0d2af265eb30f20d962cac","name":"Ziqi Wang","hidden":false},{"_id":"6a0d2af265eb30f20d962cad","name":"Zhennan Wu","hidden":false},{"_id":"6a0d2af265eb30f20d962cae","name":"Ruotong Pan","hidden":false},{"_id":"6a0d2af265eb30f20d962caf","name":"jian Liang","hidden":false},{"_id":"6a0d2af265eb30f20d962cb0","name":"Ruiming Tang","hidden":false},{"_id":"6a0d2af265eb30f20d962cb1","name":"Han Li","hidden":false}],"publishedAt":"2026-05-19T00:00:00.000Z","submittedOnDailyAt":"2026-05-20T00:00:00.000Z","title":"GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment","submittedOnDailyBy":{"_id":"61c2cf8d1172fa7969904d99","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/61c2cf8d1172fa7969904d99/R10G5h3d9Q_YQ__Hc-H4k.jpeg","isPro":false,"fullname":"suu","user":"Suu","type":"user","name":"Suu"},"summary":"We present GoLongRL, a fully open-source, capability-oriented post-training recipe for long-context reinforcement learning with verifiable rewards (RLVR). Existing long-context RL methods often treat data construction as a matter of designing increasingly complex retrieval paths, leading to homogeneous task coverage and reward formulations that inadequately reflect practical long-context requirements. Our work offers two contributions. (1) Capability-oriented data construction with full open release. We openly release a dataset of 23K RLVR samples, the complete construction pipeline, and all training code. Guided by a taxonomy of long-context capabilities, the dataset spans 9 task types, each paired with its natural evaluation metric. It comprises curated open-source samples from established corpora and synthetic samples whose QA pairs are generated from real source documents such as books, academic papers, and multi-turn dialogues. Under the same vanilla GRPO setup, our dataset alone outperforms the closed-source QwenLong-L1.5 dataset. Moreover, our Qwen3-30B-A3B model trained on this data delivers long-context performance comparable to DeepSeek-R1-0528 and Qwen3-235B-A22B-Thinking-2507, suggesting that broader coverage and greater reward diversity substantially benefit long-context capability improvement. (2) TMN-Reweight for heterogeneous multitask optimization. To address optimization challenges from heterogeneous rewards, we propose TMN-Reweight, which combines task-level mean normalization for cross-task reward scale alignment with difficulty-adaptive weighting for more reliable advantage estimation. TMN-Reweight further improves average performance over vanilla GRPO, with general capabilities preserved or improved across reported evaluations.","upvotes":44,"discussionId":"6a0d2af365eb30f20d962cb2","projectPage":"https://huggingface.co/collections/Kwai-Klear/golongrl","githubRepo":"https://github.com/xiaoxuanNLP/GoLongRL","githubRepoAddedBy":"user","ai_summary":"GoLongRL presents an open-source approach for long-context reinforcement learning with diverse reward optimization through capability-oriented data construction and TMN-Reweight methodology.","ai_keywords":["reinforcement learning","long-context","verifiable rewards","post-training recipe","RLVR","GRPO","task-level mean normalization","difficulty-adaptive weighting","advantage estimation"],"githubStars":7},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"61c2cf8d1172fa7969904d99","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/61c2cf8d1172fa7969904d99/R10G5h3d9Q_YQ__Hc-H4k.jpeg","isPro":false,"fullname":"suu","user":"Suu","type":"user"},{"_id":"65436f1c68493a7808d5bdb3","avatarUrl":"/avatars/bf528e0efd25b5e9bbb6271666a870ed.svg","isPro":false,"fullname":"dtl","user":"dtl666","type":"user"},{"_id":"664ff678c08859923b1818e2","avatarUrl":"/avatars/e28a5fd5adf171fe389d45792dd4d944.svg","isPro":false,"fullname":"oxygen","user":"oxygen0078","type":"user"},{"_id":"66cd87d0b1fe4c78fe1e3f00","avatarUrl":"/avatars/899c73530f8a4b8d8efc30f14c65ca81.svg","isPro":false,"fullname":"wzn","user":"wznwznwzn","type":"user"},{"_id":"66d9521673761d04a5856c67","avatarUrl":"/avatars/2cf4052be3a348507dcf5bac730c780b.svg","isPro":false,"fullname":"Weihan Li","user":"J1Feng","type":"user"},{"_id":"65bb205be54bc620ebc32dd6","avatarUrl":"/avatars/30288e959217d770246813aa4bb11072.svg","isPro":false,"fullname":"Xiao xinge","user":"shuiyoungwoming","type":"user"},{"_id":"62f7a3613b991129e94005da","avatarUrl":"/avatars/7ea485cb0a24bef70d24c76bc2bffa68.svg","isPro":false,"fullname":"Ruotong Pan","user":"Ruotong","type":"user"},{"_id":"64f4337d0c658d2476e6bcef","avatarUrl":"/avatars/2dd89329ae3c582ea1e1d1f0a918f195.svg","isPro":false,"fullname":"Xuan Xiao","user":"xiaoxuanzi","type":"user"},{"_id":"69364773b956a4949c79a5c5","avatarUrl":"/avatars/29dd0f107a644ea075cb94b975cfe416.svg","isPro":false,"fullname":"Liouville","user":"M-best","type":"user"},{"_id":"68249662280377699edf05ad","avatarUrl":"/avatars/bdcf0f8b7442154fc5049a1eb6f00d7c.svg","isPro":false,"fullname":"Tiehua Mei","user":"Mithas-01","type":"user"},{"_id":"6720aab6315a5b4077a64db3","avatarUrl":"/avatars/df8797737520d00f49818ff7aa2c4758.svg","isPro":false,"fullname":"chenhengrui","user":"chenhengrui","type":"user"},{"_id":"680898d9a6e089f09aa4e73b","avatarUrl":"/avatars/7ff8d7e97492512b88325c3c3ceae08f.svg","isPro":false,"fullname":"niuniukeyi","user":"niuniukeyi","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":3}">

Cite arxiv.org/abs/2605.19577 in a Space README.md to link it from this page.