Abstract
Effective information seeking in the vast and ever-growing digital landscaperequires balancing expansive search with strategic reasoning. Current largelanguage model (LLM)-based agents struggle to achieve this balance due tolimitations in search breadth and reasoning depth, where slow, serial queryingrestricts coverage of relevant sources and noisy raw inputs disrupt thecontinuity of multi-step reasoning. To address these challenges, we proposeBrowseMaster, a scalable framework built around a programmatically augmentedplanner-executor agent pair. The planner formulates and adapts searchstrategies based on task constraints, while the executor conducts efficient,targeted retrieval to supply the planner with concise, relevant evidence. Thisdivision of labor preserves coherent, long-horizon reasoning while sustainingbroad and systematic exploration, overcoming the trade-off that limits existingagents. Extensive experiments on challenging English and Chinese benchmarksshow that BrowseMaster consistently outperforms open-source and proprietarybaselines, achieving scores of 30.0 on BrowseComp-en and 46.5 on BrowseComp-zh,which demonstrates its strong capability in complex, reasoning-heavyinformation-seeking tasks at scale.