Nested Browser-Use Learning for Agentic Information Seeking

This paper addresses the limitations of current information-seeking agents, which primarily rely on API-level snippet retrieval and URL fetching, by introducing a novel framework called NestBrowse. This framework enables agents to interact with the full browser, unlocking access to richer information available through real browsing. The key innovation is a nested structure that decouples interaction control from page exploration, simplifying agentic reasoning while enabling effective deep-web information acquisition. The paper's significance lies in its potential to improve the performance of information-seeking agents on complex tasks.