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

TL;DR

Branch-and-Browse tackles the challenge of long-horizon web tasks by modeling exploration as a subtask-aware, tree-structured search that integrates reasoning, action, and contextual memory. It introduces a subtask manager, a tree exploration strategy with replay-enabled backtracking, nearest-URL state replay, background reasoning, and a page action memory to share knowledge across branches. On WebArena, it achieves a 35.8% task success rate and up to a 40.4% reduction in execution time compared to prior methods, demonstrating both depth of reasoning and exploration efficiency. This work advances practical LLM-based web agents capable of operating in dynamic, partially observable environments.

Abstract

Autonomous web agents powered by large language models (LLMs) show strong potential for performing goal-oriented tasks such as information retrieval, report generation, and online transactions. These agents mark a key step toward practical embodied reasoning in open web environments. However, existing approaches remain limited in reasoning depth and efficiency: vanilla linear methods fail at multi-step reasoning and lack effective backtracking, while other search strategies are coarse-grained and computationally costly. We introduce Branch-and-Browse, a fine-grained web agent framework that unifies structured reasoning-acting, contextual memory, and efficient execution. It (i) employs explicit subtask management with tree-structured exploration for controllable multi-branch reasoning, (ii) bootstraps exploration through efficient web state replay with background reasoning, and (iii) leverages a page action memory to share explored actions within and across sessions. On the WebArena benchmark, Branch-and-Browse achieves a task success rate of 35.8\% and reduces execution time by up to 40.4\% relative to state-of-the-art methods. These results demonstrate that Branch-and-Browse is a reliable and efficient framework for LLM-based web agents.

Figures (4)

• Figure 1: Comparison of web agent strategies. The left shows linear prompting (ReAct) with no backtracking, the middle shows inefficient tree search exploration, and the right illustrates our Branch-and-Browse framework, which enables fine-grained, memory-guided, and efficient multi-branch reasoning.
• Figure 2: Pipeline of the Branch-and-Browse framework. Given an example task zhou2023webarena, the subtask manager first decomposes the goal into three subtasks: (1) identify the page providing the product sales report, (2) filter data for the Q1 2022 period, and (3) aggregate brand sales to select the top-1 brand. Each subtask is explored through a tree-structured exploration consisting of iterative Reason–Act–Evaluation cycles. Nodes represent visited pages, and edges denote executed system actions (e.g., click, type). Depth corresponds to exploration steps, while replay links (dashed) enable efficient backtracking and reuse of prior context. As shown, successful branches (e.g., Page 8) trigger subtask_update() to advance to the next subgoal, allowing the agent to progress hierarchically through subtasks while avoiding redundant exploration.
• Figure 3: Example of reason context in branch-and-browse. During tree exploration, the agent on Page 3 reasons to execute a web interaction (browser_click), acts on the referenced element, and evaluates the outcome (score = 0), leading to Page 4. The page-level context records the task objective, progress summary, reason–action history, and page snapshot. The Page 3 Action Memory is updated by appending the new action (highlighted in blue) while marking it as irrelevant to the current goal. This cached record guides future reasoning, allowing the agent to avoid redundant exploration and efficiently backtrack during replay.
• Figure 4: Ablation results on WebArena (812 tasks): average time per successful task across all sites. Failed tasks are excluded since they often fall into repetitive exploration loops, leading to disproportionately high and uninformative time consumption.