MultiSearch: Dynamic Search Tool Orchestration for Intelligent Agents

Recent advancements in Large Reasoning Models (LRMs) have enabled agents to perform information seeking involving complex reasoning and planning. Beyond deciding what to search for, agents must also determine which search sources best satisfy their evolving information needs. However, existing agents primarily rely on single general-purpose search engines. Furthermore, selecting suitable source from multiple similar options is far more challenging than distinguishing tools with vastly different capabilities. In this paper, we introduce , a comprehensive benchmark constructed from 319 search sources with overlapping functionality and diverse content coverage. Our analysis reveals a “High Recall, Low Quality” phenomenon: naive retrieval often achieves high recall by matching surface semantics yet fails to translate this into high-quality answers. To address this, we propose a reasoning-based pipeline . Our method treats tool selection as a dynamic planning task and employs a multi-granularity reward mechanism. Extensive experiments at 4B, 8B, and 14B model scales demonstrate that our approach significantly outperforms strong baselines across various metrics. Our work lays the foundation for handling massive similar tools, paving the way for precise source selection in open-ended, multi-source deep research environments. Our code is available at: https://anonymous.4open.science/r/MultiSearchhttps://anonymous.4open.science/r/MultiSearch.