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

The article introduces Branch-and-Browse, a novel framework for autonomous web agents that enhances multi-step reasoning and efficiency by employing tree-structured exploration and contextual memory. It features explicit subtask management, efficient web state replay, and a page action memory, achieving a task success rate of 35.8% on the WebArena benchmark while reducing execution time by up to 40.4% compared to existing methods. This development is significant for practitioners as it improves the controllability and effectiveness of LLMs in performing complex web-based tasks.