We propose TreeSeeker, a framework for deep-search agents that explicitly models trial-and-error during long-horizon web research.</p>\n<p>In practice, deep search is not only a reasoning problem — it is also a search-control problem. Agents often face several plausible directions early in the search process, but only some later lead to reliable evidence. Existing systems usually follow a single evolving trajectory or a fixed execution schedule, which can cause premature commitment to weak paths.</p>\n<p>TreeSeeker addresses this by organizing search as branch-and-return exploration over tree-structured states. Our controller, TreeSearch, uses textual signals of value, uncertainty, and risk to decide whether to exploit a promising branch, explore an uncertain alternative, or prune and return from an unproductive continuation. TreeMem keeps branch-local evidence, conflicts, and failure cues so later decisions can be informed by earlier trials.</p>\n<p>Across XBench-DeepSearch, BrowseComp, and BrowseComp-ZH, TreeSeeker consistently outperforms strong open-source baselines. We hope this work contributes to building more reliable and adaptive research agents.</p>\n","updatedAt":"2026-06-12T08:39:21.486Z","author":{"_id":"6a2b64c2a696b119f6f6028b","avatarUrl":"/avatars/7d50d8b32159a34f35ddb03937d43c79.svg","fullname":"Mingzhe Ma","name":"mingzhema","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9103333950042725},"editors":["mingzhema"],"editorAvatarUrls":["/avatars/7d50d8b32159a34f35ddb03937d43c79.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.11662","authors":[{"_id":"6a2b64ca4957fcdd3aac05cd","name":"Zhuofan Shi","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05ce","user":{"_id":"6a2b64c2a696b119f6f6028b","avatarUrl":"/avatars/7d50d8b32159a34f35ddb03937d43c79.svg","isPro":false,"fullname":"Mingzhe Ma","user":"mingzhema","type":"user","name":"mingzhema"},"name":"Mingzhe Ma","status":"claimed_verified","statusLastChangedAt":"2026-06-12T06:57:59.297Z","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05cf","name":"Lu Wang","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d0","name":"Fangkai Yang","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d1","name":"Pu Zhao","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d2","name":"Yiming Guan","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d3","name":"Youling Huang","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d4","name":"Wei Zhang","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d5","name":"Qingwei Lin","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d6","name":"Dongmei Zhang","hidden":false},{"_id":"6a2b64ca4957fcdd3aac05d7","name":"Saravan Rajmohan","hidden":false}],"publishedAt":"2026-06-10T00:00:00.000Z","submittedOnDailyAt":"2026-06-12T00:00:00.000Z","title":"TreeSeeker: Tree-Structured Trial, Error, and Return in Deep Search","submittedOnDailyBy":{"_id":"6a2b64c2a696b119f6f6028b","avatarUrl":"/avatars/7d50d8b32159a34f35ddb03937d43c79.svg","isPro":false,"fullname":"Mingzhe Ma","user":"mingzhema","type":"user","name":"mingzhema"},"summary":"Deep search requires agents to answer complex questions through multi-step web search, browsing, evidence comparison, and synthesis. A central challenge is deciding how to search when several directions look plausible but only some will later lead to reliable evidence. If an agent greedily follows the current best-looking direction, it may keep extending a weak continuation. If it explores without discipline, it may waste budget on disconnected trials. We propose TreeSeeker, an inference-time framework for controlled trial-and-error in deep search. TreeSeeker organizes search as branch-and-return search over tree-structured states, where each branch is a tentative direction for a sub-goal. At each round, TreeSearch reads all sub-goal trees, identifies active goals, and uses textual UCB signals of value, uncertainty, and risk to select among exploiting a promising branch, exploring an uncertain alternative, or pruning an unproductive continuation and returning to an earlier branch point. TreeMem supports this control loop by keeping evidence, uncertainty, conflicts, progress, and failure cues attached to the branches that produced them, so trial outcomes can guide later decisions. Experiments on XBench-DeepSearch, BrowseComp, and BrowseComp-ZH show that TreeSeeker consistently outperforms strong open-source baselines, suggesting that explicit branch-and-return control complements stronger reasoning and tool execution.","upvotes":4,"discussionId":"6a2b64ca4957fcdd3aac05d8","ai_summary":"TreeSeeker is an inference-time framework that uses tree-structured search with branch-and-return control to manage exploration and exploitation in deep search tasks, improving performance through systematic trial-and-error decision making.","ai_keywords":["branch-and-return search","tree-structured states","sub-goal","textual UCB signals","value","uncertainty","risk","trial-and-error","evidence comparison","synthesis","controlled exploration"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6a2b64c2a696b119f6f6028b","avatarUrl":"/avatars/7d50d8b32159a34f35ddb03937d43c79.svg","isPro":false,"fullname":"Mingzhe Ma","user":"mingzhema","type":"user"},{"_id":"64c9ced01c25d2c581d237a8","avatarUrl":"/avatars/b4477111be53f5851c81ca3a30a784db.svg","isPro":false,"fullname":"wan","user":"Vangee","type":"user"},{"_id":"6a2b73df90fcfb6e80f04a80","avatarUrl":"/avatars/f7ac6057a2aaaf9f26797e4d882a3395.svg","isPro":false,"fullname":"Jin Rui","user":"JerryJay123","type":"user"},{"_id":"68d5d05f14dec5b7bde8885e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/MPKo6txVrsRkmwmoqizpO.png","isPro":false,"fullname":"Wei Zhang","user":"BDML-lab","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"query":{}}">
TreeSeeker: Tree-Structured Trial, Error, and Return in Deep Search
Authors: ,
,
,
,
,
,
,
,
,
Abstract
TreeSeeker is an inference-time framework that uses tree-structured search with branch-and-return control to manage exploration and exploitation in deep search tasks, improving performance through systematic trial-and-error decision making.
Deep search requires agents to answer complex questions through multi-step web search, browsing, evidence comparison, and synthesis. A central challenge is deciding how to search when several directions look plausible but only some will later lead to reliable evidence. If an agent greedily follows the current best-looking direction, it may keep extending a weak continuation. If it explores without discipline, it may waste budget on disconnected trials. We propose TreeSeeker, an inference-time framework for controlled trial-and-error in deep search. TreeSeeker organizes search as branch-and-return search over tree-structured states, where each branch is a tentative direction for a sub-goal. At each round, TreeSearch reads all sub-goal trees, identifies active goals, and uses textual UCB signals of value, uncertainty, and risk to select among exploiting a promising branch, exploring an uncertain alternative, or pruning an unproductive continuation and returning to an earlier branch point. TreeMem supports this control loop by keeping evidence, uncertainty, conflicts, progress, and failure cues attached to the branches that produced them, so trial outcomes can guide later decisions. Experiments on XBench-DeepSearch, BrowseComp, and BrowseComp-ZH show that TreeSeeker consistently outperforms strong open-source baselines, suggesting that explicit branch-and-return control complements stronger reasoning and tool execution.
Community
We propose TreeSeeker, a framework for deep-search agents that explicitly models trial-and-error during long-horizon web research.
In practice, deep search is not only a reasoning problem — it is also a search-control problem. Agents often face several plausible directions early in the search process, but only some later lead to reliable evidence. Existing systems usually follow a single evolving trajectory or a fixed execution schedule, which can cause premature commitment to weak paths.
TreeSeeker addresses this by organizing search as branch-and-return exploration over tree-structured states. Our controller, TreeSearch, uses textual signals of value, uncertainty, and risk to decide whether to exploit a promising branch, explore an uncertain alternative, or prune and return from an unproductive continuation. TreeMem keeps branch-local evidence, conflicts, and failure cues so later decisions can be informed by earlier trials.
Across XBench-DeepSearch, BrowseComp, and BrowseComp-ZH, TreeSeeker consistently outperforms strong open-source baselines. We hope this work contributes to building more reliable and adaptive research agents.
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/2606.11662 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.11662 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.11662 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.