Branch-and-Browse: Efficient and Controllable Web Exploration with Tree-Structured Reasoning and Action Memory

Researchers have introduced Branch-and-Browse, a framework for autonomous web agents that uses tree-structured reasoning to improve task success and cut execution time, according to a paper accepted at the 2026 Annual Meeting of the Association for Computational Linguistics [1][3]. The framework addresses limitations in current large language model (LLM)-based web agents, which often struggle with multi-step reasoning and lack effective backtracking [1][2]. Branch-and-Browse employs explicit subtask management with tree-structured exploration, enabling controllable multi-branch reasoning and principled backtracking [5]. It also introduces web state replay to efficiently recover the next branch for exploration, paired with background reasoning that evaluates unexplored nodes offline to prune unpromising branches and prioritize actionable steps [5]. A page action memory module records explored actions and outcomes, sharing them across branches to reduce redundancy and accelerate decision-making [5]. This module maintains structured reasoning and interaction records at the granularity of each visited page URL, allowing the agent to retrieve, summarize, and update information efficiently during both online reasoning and replay [5]. On the WebArena benchmark, Branch-and-Browse achieved a task success rate of 35.8% and reduced execution time by up to 40.4% relative to state-of-the-art methods [1][2][3]. The paper was authored by Shiqi He, Yue Cui, Xinyu Ma, Yaliang Li, Bolin Ding, and Mosharaf Chowdhury [4]. The initial submission on October 18, 2025, was 377 KB, while the revised version on June 16, 2026, expanded to 11,385 KB [1]. The work was published in the Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), spanning pages 18407–18418, and presented in San Diego, California [3]. Code is available at a public repository [5].