Real tasks can grow almost unbounded, yet a model's context is finite. We teach agentic LLMs delegation intelligence: to decompose a long-horizon task, delegate bounded subtasks to its own subagents, and integrate their condensed, evidence-grounded results, an active form of context management that lets a single model take on far more than its context alone allows.</p>\n<p>👉 Project page: <a href=\"https://search-swarm.github.io\" rel=\"nofollow\">https://search-swarm.github.io</a><br>👉 Github repo: <a href=\"https://github.com/Search-Swarm/SearchSwarm\" rel=\"nofollow\">https://github.com/Search-Swarm/SearchSwarm</a></p>\n","updatedAt":"2026-06-10T03:22:22.116Z","author":{"_id":"678532218c716c86344637e9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/678532218c716c86344637e9/UEd2G9oZkCIvaBtMJl6yX.jpeg","fullname":"Quan Chen","name":"QQqqrabit","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.9072940945625305},"editors":["QQqqrabit"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/678532218c716c86344637e9/UEd2G9oZkCIvaBtMJl6yX.jpeg"],"reactions":[],"isReport":false}},{"id":"6a294dba8dcd0c701384891b","author":{"_id":"6960eca92f7ad9b043b5cbe0","avatarUrl":"/avatars/e68dcc7fd04f143d849d40414866e633.svg","fullname":"Noah","name":"noahml","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false},"createdAt":"2026-06-10T11:42:50.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"Neat paper. The idea of using a harness to generate high-quality training data for delegation intelligence is a clever way to bypass the scarcity of this kind of logic in general text. It makes a lot of sense for long-horizon research tasks where you hit context limits quickly.\n\nHow well does the model perform when it needs to decide between delegating a subtask versus just handling it directly, and does the harness ever struggle with those edge cases?\n\nI made a podcast on it with ResearchPod, it makes it easy to get the key concepts on the go:\nhttps://researchpod.app/episode/fac229cb-493b-483d-8612-523c70cb3a5d","html":"<p>Neat paper. The idea of using a harness to generate high-quality training data for delegation intelligence is a clever way to bypass the scarcity of this kind of logic in general text. It makes a lot of sense for long-horizon research tasks where you hit context limits quickly.</p>\n<p>How well does the model perform when it needs to decide between delegating a subtask versus just handling it directly, and does the harness ever struggle with those edge cases?</p>\n<p>I made a podcast on it with ResearchPod, it makes it easy to get the key concepts on the go:<br><a href=\"https://researchpod.app/episode/fac229cb-493b-483d-8612-523c70cb3a5d\" rel=\"nofollow\">https://researchpod.app/episode/fac229cb-493b-483d-8612-523c70cb3a5d</a></p>\n","updatedAt":"2026-06-10T11:42:50.737Z","author":{"_id":"6960eca92f7ad9b043b5cbe0","avatarUrl":"/avatars/e68dcc7fd04f143d849d40414866e633.svg","fullname":"Noah","name":"noahml","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9156263470649719},"editors":["noahml"],"editorAvatarUrls":["/avatars/e68dcc7fd04f143d849d40414866e633.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.09730","authors":[{"_id":"6a27fe3c43ba362c68ab1f07","name":"Pu Ning","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f08","user":{"_id":"678532218c716c86344637e9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/678532218c716c86344637e9/UEd2G9oZkCIvaBtMJl6yX.jpeg","isPro":false,"fullname":"Quan Chen","user":"QQqqrabit","type":"user","name":"QQqqrabit"},"name":"Quan Chen","status":"claimed_verified","statusLastChangedAt":"2026-06-09T12:41:22.964Z","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f09","name":"Kun Tao","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f0a","name":"Xinyu Tang","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f0b","name":"Tianshu Wang","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f0c","name":"Qianggang Cao","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f0d","name":"Xinyu Kong","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f0e","name":"Zujie Wen","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f0f","name":"Zhiqiang Zhang","hidden":false},{"_id":"6a27fe3c43ba362c68ab1f10","name":"Jun Zhou","hidden":false}],"publishedAt":"2026-06-08T00:00:00.000Z","submittedOnDailyAt":"2026-06-10T00:00:00.000Z","title":"SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research","submittedOnDailyBy":{"_id":"678532218c716c86344637e9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/678532218c716c86344637e9/UEd2G9oZkCIvaBtMJl6yX.jpeg","isPro":false,"fullname":"Quan Chen","user":"QQqqrabit","type":"user","name":"QQqqrabit"},"summary":"Large language models are increasingly expected to handle complex, long-horizon real-world tasks whose context demands can grow without bound, yet model context windows remain inherently finite. Recent work explores a paradigm where a main agent decomposes tasks and dispatches subtasks to subagents, which execute and return only summarized results, conserving the main agent's context budget. However, performing this well requires delegation intelligence: the ability to decompose complex tasks, determine when and what to delegate, and integrate returned results into the ongoing workflow. Training data for this capability is scarce in naturally occurring text, and to our knowledge, how to synthesize such data and train models to acquire this capability remains largely unexplored in the open-source community. To bridge this gap, we present a preliminary exploration targeting deep research, a representative long-horizon agent task. Specifically, we design a harness that guides the model toward high-quality task decomposition and delegation, while constraining subagents to return results properly to support the main agent's workflow. The harness-guided trajectories naturally encode correct delegation decisions, which we use as supervised fine-tuning data to internalize delegation intelligence into model weights. Our resulting model, SearchSwarm-30B-A3B, achieves 68.1 on BrowseComp and 73.3 on BrowseComp-ZH, the best results among all models of comparable scale. We will release our harness, model weights, and training data to facilitate future research.","upvotes":43,"discussionId":"6a27fe3c43ba362c68ab1f11","projectPage":"https://search-swarm.github.io/","githubRepo":"https://github.com/Search-Swarm/SearchSwarm","githubRepoAddedBy":"user","ai_summary":"A large language model trained on synthesized delegation intelligence achieves superior performance on long-horizon research tasks through task decomposition and subagent coordination.","ai_keywords":["large language models","task decomposition","delegation intelligence","subagents","supervised fine-tuning","context windows","long-horizon tasks","SearchSwarm-30B-A3B","BrowseComp","BrowseComp-ZH"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":55,"organization":{"_id":"6a22bfdb4cfdc25bd7c9b472","name":"SearchSwarm","fullname":"SearchSwarm","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/678532218c716c86344637e9/7q_RYpvQrI7uBiS5uBdJw.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"678532218c716c86344637e9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/678532218c716c86344637e9/UEd2G9oZkCIvaBtMJl6yX.jpeg","isPro":false,"fullname":"Quan Chen","user":"QQqqrabit","type":"user"},{"_id":"6334f71031a2be3938c8b3f3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1664415607197-6334f71031a2be3938c8b3f3.jpeg","isPro":false,"fullname":"txy","user":"txy","type":"user"},{"_id":"6a28d7430930d7e8a880ec1d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a28d7430930d7e8a880ec1d/CqRBkHQTO_SCf9Yqs1PyA.jpeg","isPro":false,"fullname":"Liang Shan","user":"Twist-Shan","type":"user"},{"_id":"61ac27a7a606c4f673730377","avatarUrl":"/avatars/484fec5f93d79c49b8b88da704e29082.svg","isPro":false,"fullname":"Tianshu Wang","user":"wangtianshu","type":"user"},{"_id":"684bd729a3aa9b0f779202f0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/9yec6uHpx7pnstjs38dIY.png","isPro":false,"fullname":"RichardAloha","user":"RichardAloha","type":"user"},{"_id":"69bcf56fe26d53b997b82996","avatarUrl":"/avatars/3c648d15d9322129a168e3a135d96e05.svg","isPro":false,"fullname":"Du Fu","user":"5upern0va","type":"user"},{"_id":"6458d6067a7e192202e04c34","avatarUrl":"/avatars/2442139c9f2f7ff4718912c7e3f26bc0.svg","isPro":false,"fullname":"Qianggang Cao","user":"Xueyeeeh","type":"user"},{"_id":"68ca864a4eeb7a0096b2b1e0","avatarUrl":"/avatars/70cbfb2c1195b2c21331d64b90a3033f.svg","isPro":false,"fullname":"Junkai ZHao","user":"zajuka","type":"user"},{"_id":"674f1524b0c4efc817184e3d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/3irGg2RBUXfztsh3erZ-R.jpeg","isPro":false,"fullname":"Yang Liu","user":"Ackerlylau1999","type":"user"},{"_id":"6583e2b283a9e1460c6fb1e0","avatarUrl":"/avatars/a949165b1cec5e1d1d55f3af98182156.svg","isPro":false,"fullname":"Qianyue Hao","user":"haohao11","type":"user"},{"_id":"6638574ec78619ba63879031","avatarUrl":"/avatars/c9c95f129abfe7614493e926e7a7e971.svg","isPro":false,"fullname":"Chang Liu","user":"changliu01","type":"user"},{"_id":"6a28db15005e91d51c61316e","avatarUrl":"/avatars/52b7954bf9f5964eb63545f8744417aa.svg","isPro":false,"fullname":"Yestermorrow","user":"Yessss12138","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6a22bfdb4cfdc25bd7c9b472","name":"SearchSwarm","fullname":"SearchSwarm","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/678532218c716c86344637e9/7q_RYpvQrI7uBiS5uBdJw.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.09730.md"}">
SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research
Abstract
A large language model trained on synthesized delegation intelligence achieves superior performance on long-horizon research tasks through task decomposition and subagent coordination.
Large language models are increasingly expected to handle complex, long-horizon real-world tasks whose context demands can grow without bound, yet model context windows remain inherently finite. Recent work explores a paradigm where a main agent decomposes tasks and dispatches subtasks to subagents, which execute and return only summarized results, conserving the main agent's context budget. However, performing this well requires delegation intelligence: the ability to decompose complex tasks, determine when and what to delegate, and integrate returned results into the ongoing workflow. Training data for this capability is scarce in naturally occurring text, and to our knowledge, how to synthesize such data and train models to acquire this capability remains largely unexplored in the open-source community. To bridge this gap, we present a preliminary exploration targeting deep research, a representative long-horizon agent task. Specifically, we design a harness that guides the model toward high-quality task decomposition and delegation, while constraining subagents to return results properly to support the main agent's workflow. The harness-guided trajectories naturally encode correct delegation decisions, which we use as supervised fine-tuning data to internalize delegation intelligence into model weights. Our resulting model, SearchSwarm-30B-A3B, achieves 68.1 on BrowseComp and 73.3 on BrowseComp-ZH, the best results among all models of comparable scale. We will release our harness, model weights, and training data to facilitate future research.
Community
Real tasks can grow almost unbounded, yet a model's context is finite. We teach agentic LLMs delegation intelligence: to decompose a long-horizon task, delegate bounded subtasks to its own subagents, and integrate their condensed, evidence-grounded results, an active form of context management that lets a single model take on far more than its context alone allows.
👉 Project page: https://search-swarm.github.io
👉 Github repo: https://github.com/Search-Swarm/SearchSwarm
Neat paper. The idea of using a harness to generate high-quality training data for delegation intelligence is a clever way to bypass the scarcity of this kind of logic in general text. It makes a lot of sense for long-horizon research tasks where you hit context limits quickly.
How well does the model perform when it needs to decide between delegating a subtask versus just handling it directly, and does the harness ever struggle with those edge cases?
I made a podcast on it with ResearchPod, it makes it easy to get the key concepts on the go:
https://researchpod.app/episode/fac229cb-493b-483d-8612-523c70cb3a5d
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/2606.09730 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.09730 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.09730 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.