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

TreeSeeker is a new framework for deep search that utilizes a tree-structured approach to manage trial-and-error during inference. It employs a controlled exploration strategy, leveraging textual Upper Confidence Bound (UCB) signals to navigate between promising branches and prune unproductive paths, while TreeMem maintains contextual information throughout the search process. Experimental results indicate that TreeSeeker outperforms existing open-source baselines on benchmarks like XBench-DeepSearch and BrowseComp, highlighting its potential for enhancing decision-making in complex AI search tasks.