Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision. Knowledge distillation can supply denser guidance, and advanced proprietary models with their strong reasoning capabilities are promising teachers. While distilling from proprietary models can densify this supervisory signal, conventional logit-matching is precluded by hidden logits and mismatched tokenizers, whereas raw natural language trajectory imitation transfers superficial stylistic artifacts rather than core reasoning competence. To address the heterogeneous distillation problem and bridge the distribution gap, we propose Multi-Agent Protocol Distillation (MAPD), a joint distillation and RL framework uses a structured, style-normalized protocol as an intermediate representation. An offline multi-agent system (MAS) decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the resulting exploration trace into a JSON protocol containing the task type, reasoning plan, and extractive grounding facts. During training, the protocol is provided only to a privileged branch of the student policy, whose token distributions furnish a dense distillation signal alongside the sparse RL objective. Extensive evaluations across seven QA benchmarks demonstrate that MAPD consistently outperforms competitive distillation and RL, achieving average success rates of 39.4% on Qwen3-1.7B and 44.4% on Qwen3-4B. Crucially, the framework generalizes robustly across diverse proprietary teachers while effectively mitigating the student policy from style drift and verbosity degeneration.</p>\n","updatedAt":"2026-07-28T02:54:18.883Z","author":{"_id":"65745569839aa08899ea5d27","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/4X8waDwiphbfKZySrYlFy.jpeg","fullname":"Kailin Jiang","name":"kailinjiang","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":5,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8796855807304382},"editors":["kailinjiang"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/4X8waDwiphbfKZySrYlFy.jpeg"],"reactions":[{"reaction":"👀","users":["jvan77"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.24280","authors":[{"_id":"6a68189173f69d5af2bec617","name":"Junlin Liu","hidden":false},{"_id":"6a68189173f69d5af2bec618","name":"Jiangwang Chen","hidden":false},{"_id":"6a68189173f69d5af2bec619","name":"Zixin Song","hidden":false},{"_id":"6a68189173f69d5af2bec61a","name":"Shuaiyu Zhou","hidden":false},{"_id":"6a68189173f69d5af2bec61b","name":"Chunji Lv","hidden":false},{"_id":"6a68189173f69d5af2bec61c","name":"Hank Wu","hidden":false},{"_id":"6a68189173f69d5af2bec61d","user":{"_id":"65745569839aa08899ea5d27","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/4X8waDwiphbfKZySrYlFy.jpeg","isPro":false,"fullname":"Kailin Jiang","user":"kailinjiang","type":"user","name":"kailinjiang"},"name":"Kailin Jiang","status":"claimed_verified","statusLastChangedAt":"2026-07-28T08:45:04.578Z","hidden":false},{"_id":"6a68189173f69d5af2bec61e","user":{"_id":"6747de57f8cab58c22ec94a2","avatarUrl":"/avatars/5bae0341862fac24564781c0fa32aac5.svg","isPro":false,"fullname":"Jinyang Wu","user":"Jinyang23","type":"user","name":"Jinyang23"},"name":"Jinyang Wu","status":"claimed_verified","statusLastChangedAt":"2026-07-28T08:45:04.583Z","hidden":false},{"_id":"6a68189173f69d5af2bec61f","name":"Bohan Yu","hidden":false},{"_id":"6a68189173f69d5af2bec620","name":"Chenxi Zhou","hidden":false}],"publishedAt":"2026-07-27T00:00:00.000Z","submittedOnDailyAt":"2026-07-28T00:00:00.000Z","title":"From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search","submittedOnDailyBy":{"_id":"65745569839aa08899ea5d27","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/4X8waDwiphbfKZySrYlFy.jpeg","isPro":false,"fullname":"Kailin Jiang","user":"kailinjiang","type":"user","name":"kailinjiang"},"summary":"Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision. Knowledge distillation can supply denser guidance, and advanced proprietary models with their strong reasoning capabilities are promising teachers. While distilling from proprietary models can densify this supervisory signal, conventional logit-matching is precluded by hidden logits and mismatched tokenizers, whereas raw natural language trajectory imitation transfers superficial stylistic artifacts rather than core reasoning competence. To address the heterogeneous distillation problem and bridge the distribution gap, we propose Multi-Agent Protocol Distillation (MAPD), a joint distillation and RL framework uses a structured, style-normalized protocol as an intermediate representation. An offline multi-agent system (MAS) decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the resulting exploration trace into a JSON protocol containing the task type, reasoning plan, and extractive grounding facts. During training, the protocol is provided only to a privileged branch of the student policy, whose token distributions furnish a dense distillation signal alongside the sparse RL objective. Extensive evaluations across seven QA benchmarks demonstrate that MAPD consistently outperforms competitive distillation and RL, achieving average success rates of 39.4\\% on Qwen3-1.7B and 44.4\\% on Qwen3-4B. Crucially, the framework generalizes robustly across diverse proprietary teachers while effectively mitigating the student policy from style drift and verbosity degeneration.","upvotes":64,"discussionId":"6a68189173f69d5af2bec621"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"69045a2e9c3d523381fcf489","avatarUrl":"/avatars/8b1982e043a7ab28006b707dfff6c5dc.svg","isPro":false,"fullname":"JunlinLiu","user":"AaronLiu0702","type":"user"},{"_id":"65745569839aa08899ea5d27","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/4X8waDwiphbfKZySrYlFy.jpeg","isPro":false,"fullname":"Kailin Jiang","user":"kailinjiang","type":"user"},{"_id":"6a53973b0fe5d0e37b02f80d","avatarUrl":"/avatars/498a1d273d9459f4c6d0d9e37ff4d286.svg","isPro":false,"fullname":"doppel chen","user":"doppelgangerX","type":"user"},{"_id":"6581b6038ab7e9a61656321b","avatarUrl":"/avatars/adab1b1704a11964f395061ded6d1472.svg","isPro":false,"fullname":"yapeng mi","user":"MickeyMouse2024","type":"user"},{"_id":"65372bae0d973d3fee4131c7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/OeBjr1hlUINmjUJ8iu42I.png","isPro":false,"fullname":"BadCat","user":"Foresta","type":"user"},{"_id":"676a1264257409ac8a976587","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/m9ZDoBW_Do8YL1kRGNn_J.png","isPro":false,"fullname":"ZelateCalcite","user":"Zelate","type":"user"},{"_id":"6747de57f8cab58c22ec94a2","avatarUrl":"/avatars/5bae0341862fac24564781c0fa32aac5.svg","isPro":false,"fullname":"Jinyang Wu","user":"Jinyang23","type":"user"},{"_id":"68d35c61a4faedb167e99546","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/U2552CQOGaymVamgxcmfM.png","isPro":false,"fullname":"chunji lv","user":"yaya234","type":"user"},{"_id":"652902d9d0c969f26518013c","avatarUrl":"/avatars/2b2c49028b0ee613c5c378d1c96c3649.svg","isPro":false,"fullname":"Jianzhe Ma","user":"majianz","type":"user"},{"_id":"64c4accbb496b4e176786171","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64c4accbb496b4e176786171/eyh6ulc4fufpGmf_DO2K3.jpeg","isPro":false,"fullname":"Molly Sophia","user":"mollysama","type":"user"},{"_id":"66eb99b061228b02f84dbf70","avatarUrl":"/avatars/94e94a457168428543ac29f2bab7b499.svg","isPro":false,"fullname":"Zrp Zhang","user":"RuipengZhang","type":"user"},{"_id":"68bdae76167c90e2e86584cb","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/WfVyTk2onc6jMB70ROVOp.png","isPro":false,"fullname":"victerrrrr","user":"victerrrrr","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.24280.md","query":{}}">
From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search
Abstract
Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision. Knowledge distillation can supply denser guidance, and advanced proprietary models with their strong reasoning capabilities are promising teachers. While distilling from proprietary models can densify this supervisory signal, conventional logit-matching is precluded by hidden logits and mismatched tokenizers, whereas raw natural language trajectory imitation transfers superficial stylistic artifacts rather than core reasoning competence. To address the heterogeneous distillation problem and bridge the distribution gap, we propose Multi-Agent Protocol Distillation (MAPD), a joint distillation and RL framework uses a structured, style-normalized protocol as an intermediate representation. An offline multi-agent system (MAS) decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the resulting exploration trace into a JSON protocol containing the task type, reasoning plan, and extractive grounding facts. During training, the protocol is provided only to a privileged branch of the student policy, whose token distributions furnish a dense distillation signal alongside the sparse RL objective. Extensive evaluations across seven QA benchmarks demonstrate that MAPD consistently outperforms competitive distillation and RL, achieving average success rates of 39.4\% on Qwen3-1.7B and 44.4\% on Qwen3-4B. Crucially, the framework generalizes robustly across diverse proprietary teachers while effectively mitigating the student policy from style drift and verbosity degeneration.
Community
Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision. Knowledge distillation can supply denser guidance, and advanced proprietary models with their strong reasoning capabilities are promising teachers. While distilling from proprietary models can densify this supervisory signal, conventional logit-matching is precluded by hidden logits and mismatched tokenizers, whereas raw natural language trajectory imitation transfers superficial stylistic artifacts rather than core reasoning competence. To address the heterogeneous distillation problem and bridge the distribution gap, we propose Multi-Agent Protocol Distillation (MAPD), a joint distillation and RL framework uses a structured, style-normalized protocol as an intermediate representation. An offline multi-agent system (MAS) decomposes each query, retrieves supporting evidence, repairs failed searches, and converts the resulting exploration trace into a JSON protocol containing the task type, reasoning plan, and extractive grounding facts. During training, the protocol is provided only to a privileged branch of the student policy, whose token distributions furnish a dense distillation signal alongside the sparse RL objective. Extensive evaluations across seven QA benchmarks demonstrate that MAPD consistently outperforms competitive distillation and RL, achieving average success rates of 39.4% on Qwen3-1.7B and 44.4% on Qwen3-4B. Crucially, the framework generalizes robustly across diverse proprietary teachers while effectively mitigating the student policy from style drift and verbosity degeneration.
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/2607.24280 in a model README.md to link it from this page.
Cite arxiv.org/abs/2607.24280 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2607.24280 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.