Zhuofan Shi, Mingzhe Ma, Lu Wang, Fangkai Yang, Pu Zhao, Yiming Guan, Youling Huang, Wei Zhang, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan
4 min
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.
Alex: Failure cues?
Sam: Little markers that say, in effect, "this path was tried and didn't work, and here's why." So when the agent is deciding whether to revisit a similar direction later, it has a concrete reason not to. It's not just forgetting dead ends—it's learning from them.
Alex: That's the part that seems most significant to me. It's not just managing multiple paths—it's building a record of what didn't work and using that to make better decisions going forward.
Sam: That's the core insight. And it keeps the whole process disciplined. At each decision round, the agent surveys all its active sub-goals and makes one clear choice for each: keep going, try something new, or cut it. It's not trying to rank every possible action in the universe—just making a focused call on each live branch.
Alex: Does that add up to a meaningful improvement in practice?
Sam: The paper reports that this structured approach—the branching, the failure memory, the explicit pruning—consistently outperforms standard baseline models on multi-step research tasks. The authors frame the explicit "prune and return" rule as a meaningful step forward for this type of agent.
Alex: Are there limits to what it can do?
Sam: A couple worth noting. Right now, TreeSeeker is strictly text-based—it can't process images or other media. And because the agent pauses to summarize and make strategic decisions at each branch point, there's a delay. You get higher-quality research, but it takes longer than a simpler agent that just barrels forward.
Alex: So there's a trade-off between depth and speed.
Sam: Precisely. The extra thinking time is the cost of that higher accuracy. The next challenge for the field is making that decision-making more efficient—so you don't have to choose between the two.
Alex: That's a clear and honest picture of where things stand. A more disciplined, self-correcting approach to research—with real limitations still to work through. Thanks for listening to ResearchPod.