🚀 We introduce <strong>DART-SD</strong>, a topology-aware self-distillation framework for multi-turn tool-calling agents.</p>\n<p>Existing SFT and RL methods typically treat tool-use trajectories as linear sequences, which can incorrectly penalize valid alternative exploration. DART-SD instead models tool interactions as an <strong>Interaction-State Transition Graph (ISTG)</strong>, capturing the inherent <em>diamond topology</em> induced by order-independent sub-goals.</p>\n<p>Based on this structure, DART-SD identifies the <strong>Critical Topological Breakpoint (CTB)</strong> of failed trajectories, retrieves success-supported recovery paths, and applies <strong>localized supervision only after the breakpoint</strong>, preserving already-correct reasoning prefixes.</p>\n<p>Across five tool-use benchmarks and two model scales, DART-SD consistently outperforms strong SFT and RL baselines. Notably, the Qwen3-8B student even surpasses its much larger teacher on several benchmarks.</p>\n<p>💡 Rather than forcing agents to imitate entire trajectories, DART-SD learns <strong>where the reasoning actually goes wrong — and only fixes from there</strong>.</p>\n","updatedAt":"2026-08-31T07:00:39.322Z","author":{"_id":"6954da80bc38ca0baf130a68","avatarUrl":"/avatars/e43cd2bd28c591dc0deabd808dc27e79.svg","fullname":"Hangrui Xu","name":"Hangrui-Xu","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8775051236152649},"editors":["Hangrui-Xu"],"editorAvatarUrls":["/avatars/e43cd2bd28c591dc0deabd808dc27e79.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.18524","authors":[{"_id":"6a952572073195fee51572ed","user":{"_id":"6954da80bc38ca0baf130a68","avatarUrl":"/avatars/e43cd2bd28c591dc0deabd808dc27e79.svg","isPro":false,"fullname":"Hangrui Xu","user":"Hangrui-Xu","type":"user","name":"Hangrui-Xu"},"name":"Hangrui Xu","status":"claimed_verified","statusLastChangedAt":"2026-08-31T08:37:51.524Z","hidden":false},{"_id":"6a952572073195fee51572ee","name":"Jiarui Wang","hidden":false},{"_id":"6a952572073195fee51572ef","name":"Yang Yang","hidden":false},{"_id":"6a952572073195fee51572f0","name":"Chuanbo Zhu","hidden":false},{"_id":"6a952572073195fee51572f1","name":"Fangda Chen","hidden":false},{"_id":"6a952572073195fee51572f2","name":"Ziqi Wu","hidden":false},{"_id":"6a952572073195fee51572f3","name":"Jingming Cai","hidden":false},{"_id":"6a952572073195fee51572f4","name":"Yan Song","hidden":false}],"publishedAt":"2026-08-19T00:00:00.000Z","submittedOnDailyAt":"2026-08-31T00:00:00.000Z","title":"DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents","submittedOnDailyBy":{"_id":"6954da80bc38ca0baf130a68","avatarUrl":"/avatars/e43cd2bd28c591dc0deabd808dc27e79.svg","isPro":false,"fullname":"Hangrui Xu","user":"Hangrui-Xu","type":"user","name":"Hangrui-Xu"},"summary":"Equipping Large Language Models (LLMs) with multi-turn tool-calling capabilities is essential for building autonomous agents. However, progress is fundamentally limited by the reliance on full-length trajectory imitation. For tasks involving multiple order-independent sub-goals, the optimal solution space forms a vast combinatorial diamond lattice. Forcing this rich topology into monolithic trajectories causes a severe topological collapse, indiscriminately penalizing valid alternative explorations and severely degrading policy diversity. To address this, we propose DART-SD (Diamond-topology Aware Retrieval and Tuning for Self-Distillation), a novel framework that shifts the paradigm from global forcing to topology-guided localized correction. DART-SD first models the execution process as a converging Interaction-State Transition Graph (ISTG), faithfully capturing the inherent diamond topology of successful and failed exploratory paths. During autonomous rollouts, the framework identifies the Critical Topological Breakpoint (CTB) and retrieves success-supported recovery references. Finally, we introduce a progressive self-distillation paradigm through CTB-guided localized supervision, ensuring that the training loss is calculated exclusively on the generated recovery steps while strictly protecting the valid reasoning prefix from destructive gradient updates. Experiments on complex multi-turn tool-calling benchmarks demonstrate that DART-SD significantly outperforms traditional full-trajectory baselines.","upvotes":59,"discussionId":"6a952572073195fee51572f5","ai_summary":"DART-SD improves multi-turn tool-calling agents by modeling execution as a diamond-topology graph, identifying critical failure points, and applying localized self-distillation to preserve valid reasoning while correcting errors.","ai_keywords":["DART-SD","Interaction-State Transition Graph","ISTG","Critical Topological Breakpoint","CTB","self-distillation","diamond topology","multi-turn tool-calling"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"653b817d32c97d0655575872","name":"ByteDance","fullname":"ByteDance","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6535c9e88bde2fae19b6fb25/0clr54wj5Ly-RkYU9OXPp.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6954da80bc38ca0baf130a68","avatarUrl":"/avatars/e43cd2bd28c591dc0deabd808dc27e79.svg","isPro":false,"fullname":"Hangrui Xu","user":"Hangrui-Xu","type":"user"},{"_id":"6a86bfef2eefbc90b6de3b36","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a86bfef2eefbc90b6de3b36/moBBVelLm6ZA9piJlDR7D.jpeg","isPro":false,"fullname":"Hyejin V. Kim","user":"hyejinkimva","type":"user"},{"_id":"6a87f770d1a23466e82827c3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a87f770d1a23466e82827c3/HYiXTea9w2yAVKRMBwZVE.jpeg","isPro":false,"fullname":"Artem G. Lebedev","user":"artemlxdp","type":"user"},{"_id":"6a8d7d47843b8411841b63b4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a8d7d47843b8411841b63b4/mKco6fHBBgW2PAJEiJMFY.jpeg","isPro":false,"fullname":"Laura","user":"oliveix2003","type":"user"},{"_id":"6a8d55acb955e313e36d7ff7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a8d55acb955e313e36d7ff7/eXLF2h5XXxGjjZKFuvbP3.jpeg","isPro":false,"fullname":"नेहा पटेल","user":"ashishguptagos","type":"user"},{"_id":"6a814baa1c2a53b66ef18585","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a814baa1c2a53b66ef18585/LeG32PrHqVavxEZ7m7g_s.jpeg","isPro":false,"fullname":"Vincent Lee","user":"RiccardoBarbieri","type":"user"},{"_id":"6a7d391dbd1db6a4382e21ae","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a7d391dbd1db6a4382e21ae/thJav6-n7ZT0tLHStBv_j.jpeg","isPro":false,"fullname":"Diego Sánchez-Almeida","user":"joseph-smith","type":"user"},{"_id":"6a835ab090bdc422718cf93c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a835ab090bdc422718cf93c/hNgCW0m3psrnyjOrhInU-.jpeg","isPro":false,"fullname":"Sophie P. Smith","user":"sophiesmith","type":"user"},{"_id":"6a84a1dcd143207e489b962c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a84a1dcd143207e489b962c/6nMk-homRQJsfdRfWiHbo.jpeg","isPro":false,"fullname":"李梓涵","user":"weihan85","type":"user"},{"_id":"6a7f03b81456db8389c50a58","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a7f03b81456db8389c50a58/aE5PHDgDFZfewBxyzppYz.jpeg","isPro":false,"fullname":"Cindy Kim","user":"cindykim","type":"user"},{"_id":"6a82147dc328fa48187cf64d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a82147dc328fa48187cf64d/mWxFZ8w2K7VvIHNLlxhnI.jpeg","isPro":false,"fullname":"M. Garcia","user":"melissa-garcia","type":"user"},{"_id":"6a7e1c2606f88595d94b484f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a7e1c2606f88595d94b484f/hhAuJzDSJE4YVnmxhSrNc.jpeg","isPro":false,"fullname":"Leo Thomas-Wilson","user":"aalqahtaninoura","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":1,"organization":{"_id":"653b817d32c97d0655575872","name":"ByteDance","fullname":"ByteDance","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6535c9e88bde2fae19b6fb25/0clr54wj5Ly-RkYU9OXPp.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.18524.md","query":{}}">
DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents
Abstract
DART-SD improves multi-turn tool-calling agents by modeling execution as a diamond-topology graph, identifying critical failure points, and applying localized self-distillation to preserve valid reasoning while correcting errors.
Equipping Large Language Models (LLMs) with multi-turn tool-calling capabilities is essential for building autonomous agents. However, progress is fundamentally limited by the reliance on full-length trajectory imitation. For tasks involving multiple order-independent sub-goals, the optimal solution space forms a vast combinatorial diamond lattice. Forcing this rich topology into monolithic trajectories causes a severe topological collapse, indiscriminately penalizing valid alternative explorations and severely degrading policy diversity. To address this, we propose DART-SD (Diamond-topology Aware Retrieval and Tuning for Self-Distillation), a novel framework that shifts the paradigm from global forcing to topology-guided localized correction. DART-SD first models the execution process as a converging Interaction-State Transition Graph (ISTG), faithfully capturing the inherent diamond topology of successful and failed exploratory paths. During autonomous rollouts, the framework identifies the Critical Topological Breakpoint (CTB) and retrieves success-supported recovery references. Finally, we introduce a progressive self-distillation paradigm through CTB-guided localized supervision, ensuring that the training loss is calculated exclusively on the generated recovery steps while strictly protecting the valid reasoning prefix from destructive gradient updates. Experiments on complex multi-turn tool-calling benchmarks demonstrate that DART-SD significantly outperforms traditional full-trajectory baselines.
Community
🚀 We introduce DART-SD, a topology-aware self-distillation framework for multi-turn tool-calling agents.
Existing SFT and RL methods typically treat tool-use trajectories as linear sequences, which can incorrectly penalize valid alternative exploration. DART-SD instead models tool interactions as an Interaction-State Transition Graph (ISTG), capturing the inherent diamond topology induced by order-independent sub-goals.
Based on this structure, DART-SD identifies the Critical Topological Breakpoint (CTB) of failed trajectories, retrieves success-supported recovery paths, and applies localized supervision only after the breakpoint, preserving already-correct reasoning prefixes.
Across five tool-use benchmarks and two model scales, DART-SD consistently outperforms strong SFT and RL baselines. Notably, the Qwen3-8B student even surpasses its much larger teacher on several benchmarks.
💡 Rather than forcing agents to imitate entire trajectories, DART-SD learns where the reasoning actually goes wrong — and only fixes from there.
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/2608.18524 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.18524 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.18524 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.