TreeSeeker: Tree-Structured Trial, Error, and Return in Deep Search

arXiv cs.AI Papers
deep-search tree-structured trial-and-error branch-and-return web-search agent

Summary

TreeSeeker is an inference-time framework that organizes deep search as branch-and-return over tree-structured states, using textual UCB signals to balance exploitation, exploration, and pruning. It outperforms strong baselines on deep search benchmarks, showing that explicit branch-and-return control improves multi-step web search.

arXiv:2606.11662v1 Announce Type: new Abstract: 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.
Original Article
View Cached Full Text

Cached at: 06/11/26, 01:48 PM 

# TreeSeeker: Tree-Structured Trial, Error, and Return in Deep Search
Source: [https://arxiv.org/abs/2606.11662](https://arxiv.org/abs/2606.11662)
Authors:[Zhuofan Shi](https://arxiv.org/search/cs?searchtype=author&query=Shi,+Z),[Mingzhe Ma](https://arxiv.org/search/cs?searchtype=author&query=Ma,+M),[Lu Wang](https://arxiv.org/search/cs?searchtype=author&query=Wang,+L),[Fangkai Yang](https://arxiv.org/search/cs?searchtype=author&query=Yang,+F),[Pu Zhao](https://arxiv.org/search/cs?searchtype=author&query=Zhao,+P),[Yiming Guan](https://arxiv.org/search/cs?searchtype=author&query=Guan,+Y),[Youling Huang](https://arxiv.org/search/cs?searchtype=author&query=Huang,+Y),[Wei Zhang](https://arxiv.org/search/cs?searchtype=author&query=Zhang,+W),[Qingwei Lin](https://arxiv.org/search/cs?searchtype=author&query=Lin,+Q),[Dongmei Zhang](https://arxiv.org/search/cs?searchtype=author&query=Zhang,+D),[Saravan Rajmohan](https://arxiv.org/search/cs?searchtype=author&query=Rajmohan,+S)

[View PDF](https://arxiv.org/pdf/2606.11662)

> Abstract: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\.

## Submission history

From: Mingzhe Ma \[[view email](https://arxiv.org/show-email/758d8abd/2606.11662)\] **\[v1\]**Wed, 10 Jun 2026 05:10:19 UTC \(951 KB\)

Similar Articles

Structure-Induced Information for Rerooting Levin Tree Search

arXiv cs.AI

This paper proposes three rerooter designs for Levin Tree Search that leverage state-space structure and learned heuristics to improve search efficiency without explicit subgoal generation, achieving state-of-the-art online training efficiency.

Arbor: Tree Search as a Cognition Layer for Autonomous Agents

arXiv cs.AI

Arbor introduces structured tree search as a cognition layer for autonomous agents, enabling multi-day, full-stack LLM inference optimization with up to 193% throughput-latency improvement over vendor baselines through a checks-and-balances multi-agent architecture.

ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

arXiv cs.AI

This paper proposes Answer-Backtracked Credit Assignment (ABC), a framework that converts sparse trajectory-level outcomes into dense step-level supervision for training long-horizon search agents. The resulting ABSeeker model, built on Qwen3.5-4B, achieves strong results on BrowseComp benchmarks, outperforming same-scale agents and matching larger models.

SearchEyes: Towards Frontier Multimodal Deep Search Intelligence via Search World Simulation

arXiv cs.AI

SearchEyes uses a typed knowledge graph as the backbone of a simulated search world to unify training data, search environment, and reward signals. It proposes Perception-Knowledge Chains (PKC) for multi-hop path sampling and Hop-Anchored Policy Optimization (HaPO) for step-level credit assignment, achieving state-of-the-art performance on multimodal knowledge-intensive benchmarks.