TAC studies general reasoning through the lens of transferability: instead of asking whether post-training improves performance on its source domain, we ask how well the learned behavior transfers across held-out domains.</p>\n<p>Across 14 benchmarks in 6 domains and two backbones, TAC improves macro-average accuracy and reveals a surprising pattern: math, often treated as a central RLVR domain, is among the least transferable.<br><a href=\"https://cdn-uploads.huggingface.co/production/uploads/6390438588139d16aae471b7/C0t2pml8APl3z6kW34Orf.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/6390438588139d16aae471b7/C0t2pml8APl3z6kW34Orf.png\" alt=\"overview\"></a></p>\n","updatedAt":"2026-07-03T07:02:03.948Z","author":{"_id":"6390438588139d16aae471b7","avatarUrl":"/avatars/18ffd5f4a5afece842d4ac89e1d1596e.svg","fullname":"Yang Yongjin","name":"Dragonjinny","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8537046313285828},"editors":["Dragonjinny"],"editorAvatarUrls":["/avatars/18ffd5f4a5afece842d4ac89e1d1596e.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.25178","authors":[{"_id":"6a43e42541f04ae4d7ad9556","user":{"_id":"6390438588139d16aae471b7","avatarUrl":"/avatars/18ffd5f4a5afece842d4ac89e1d1596e.svg","isPro":false,"fullname":"Yang Yongjin","user":"Dragonjinny","type":"user","name":"Dragonjinny"},"name":"Yongjin Yang","status":"claimed_verified","statusLastChangedAt":"2026-07-01T08:45:11.698Z","hidden":false},{"_id":"6a43e42541f04ae4d7ad9557","name":"Jiarui Liu","hidden":false},{"_id":"6a43e42541f04ae4d7ad9558","name":"Yinghui He","hidden":false},{"_id":"6a43e42541f04ae4d7ad9559","name":"Lechen Zhang","hidden":false},{"_id":"6a43e42541f04ae4d7ad955a","name":"Bernhard Schölkopf","hidden":false},{"_id":"6a43e42541f04ae4d7ad955b","name":"Zhijing Jin","hidden":false}],"publishedAt":"2026-06-27T00:00:00.000Z","submittedOnDailyAt":"2026-07-03T00:00:00.000Z","title":"Transferability for General Reasoning: An Automated Curriculum for Multi-Domain RLVR","submittedOnDailyBy":{"_id":"6390438588139d16aae471b7","avatarUrl":"/avatars/18ffd5f4a5afece842d4ac89e1d1596e.svg","isPro":false,"fullname":"Yang Yongjin","user":"Dragonjinny","type":"user","name":"Dragonjinny"},"summary":"Reinforcement learning with verifiable rewards (RLVR) has been extended from single-domain training to multi-domain reasoning suites spanning mathematics, programming, and science. However, the training curriculum (how often each domain is sampled) is typically fixed or hand-tuned, even though reasoning skills transfer unevenly across domains. Existing learnability-based curricula adapt to where the policy is currently improving, but are blind to whether a gradient step on the selected domain benefits the remaining domains. In this paper, we propose Transfer-Aware Curriculum (TAC), a bandit-style online curriculum that prioritizes domains whose updates broadly benefit the rest of the training suite. TAC repurposes signals already produced by RL training: per-domain advantages capture local learnability, and projected gradients, taken from the GRPO step being computed, estimate cross-domain transferability via gradient-geometry alignment, at negligible cost (<1% wall-clock overhead). Across a six-domain reasoning suite, TAC achieves the best macro-averaged accuracy on both Qwen3-1.7B and Llama3.2-3B, outperforming proportional random sampling, a hand-designed schedule, and a learnability-only bandit, and improving over the last of these by up to 2.8 points (10% relative). Ablations show performance degrades sharply when the transferability term is removed, and TAC remains robust on imbalanced training mixtures where learnability-only curricula over-commit to dominant domains. Our findings establish cross-domain transferability as a key signal for curriculum design in multi-domain RLVR.","upvotes":2,"discussionId":"6a43e42541f04ae4d7ad955c","githubRepo":"https://github.com/YangYongJin/transfer-aware-curriculum","githubRepoAddedBy":"user","ai_summary":"Transfer-Aware Curriculum (TAC) improves multi-domain reinforcement learning by prioritizing domains that provide broad benefits to other domains, using gradient-geometry alignment to estimate cross-domain transferability.","ai_keywords":["reinforcement learning","verifiable rewards","multi-domain reasoning","curriculum learning","bandit-style online curriculum","transferability","gradient-geometry alignment","GRPO","macro-averaged accuracy"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":2,"organization":{"_id":"62c5000b4d3cf26ce7c62822","name":"uoft","fullname":"University of Toronto","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1657077766523-62c4ff85cb7033fd49b7a559.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6390438588139d16aae471b7","avatarUrl":"/avatars/18ffd5f4a5afece842d4ac89e1d1596e.svg","isPro":false,"fullname":"Yang Yongjin","user":"Dragonjinny","type":"user"},{"_id":"69cd3d87729d8ce47bc735b8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/sP9Sb541Ez_DjDxUC4vg1.png","isPro":false,"fullname":"周 佳慧","user":"jacksonba","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"62c5000b4d3cf26ce7c62822","name":"uoft","fullname":"University of Toronto","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1657077766523-62c4ff85cb7033fd49b7a559.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.25178.md","query":{}}">
Transferability for General Reasoning: An Automated Curriculum for Multi-Domain RLVR
Abstract
Transfer-Aware Curriculum (TAC) improves multi-domain reinforcement learning by prioritizing domains that provide broad benefits to other domains, using gradient-geometry alignment to estimate cross-domain transferability.
Reinforcement learning with verifiable rewards (RLVR) has been extended from single-domain training to multi-domain reasoning suites spanning mathematics, programming, and science. However, the training curriculum (how often each domain is sampled) is typically fixed or hand-tuned, even though reasoning skills transfer unevenly across domains. Existing learnability-based curricula adapt to where the policy is currently improving, but are blind to whether a gradient step on the selected domain benefits the remaining domains. In this paper, we propose Transfer-Aware Curriculum (TAC), a bandit-style online curriculum that prioritizes domains whose updates broadly benefit the rest of the training suite. TAC repurposes signals already produced by RL training: per-domain advantages capture local learnability, and projected gradients, taken from the GRPO step being computed, estimate cross-domain transferability via gradient-geometry alignment, at negligible cost (<1% wall-clock overhead). Across a six-domain reasoning suite, TAC achieves the best macro-averaged accuracy on both Qwen3-1.7B and Llama3.2-3B, outperforming proportional random sampling, a hand-designed schedule, and a learnability-only bandit, and improving over the last of these by up to 2.8 points (10% relative). Ablations show performance degrades sharply when the transferability term is removed, and TAC remains robust on imbalanced training mixtures where learnability-only curricula over-commit to dominant domains. Our findings establish cross-domain transferability as a key signal for curriculum design in multi-domain RLVR.
Community
TAC studies general reasoning through the lens of transferability: instead of asking whether post-training improves performance on its source domain, we ask how well the learned behavior transfers across held-out domains.
Across 14 benchmarks in 6 domains and two backbones, TAC improves macro-average accuracy and reveals a surprising pattern: math, often treated as a central RLVR domain, is among the least transferable.
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/2606.25178 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.25178 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.25178 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.