Overview
Field: Machine Learning Authors: Yuyao Zhang, Junjie Gao, Zhengxian Wu, Jiaming Fan, Jin Zhang, Shihan Ma, Yao Yao, Weiran Qi, Chuyan Jin, Guiyu Ma, Xingzhong Xu, Kai Yang, Ji-Rong Wen, Zhicheng Dou Published: 2026-07-16 arXiv: 2607.15257
Summary
Recent advances in Tool-Integrated Large Language Models have made web search a core capability of information-seeking agents. However, as interaction histories grow, agents increasingly struggle to track task progress. When search attempts fail to yield useful evidence, current single- and multi-agent systems can become trapped in repetitive loops, wasting search budgets and ultimately compromising the quality and completeness of the final output.
SearchOS is a system-level multi-agent framework that turns fragile, implicit search progress into explicit, persistent, and shared state. Its key components:
- Task formulation: Open-domain information seeking is framed as relational schema completion with grounded citations — agents discover entities, populate attributes across linked tables, and anchor each value to source evidence.
- Search-Oriented Context Management (SOCM): Externalizes evolving state into frontier tasks, an evidence graph, a coverage map, and failure memory.
- Pipelined parallel scheduling: Overlaps sub-agent execution and continuously backfills freed slots with tasks targeting unresolved coverage gaps, improving utilization and throughput.
- Search tool middleware harness: Intercepts model-tool interactions to record citation-grounded evidence and react to stalls or budget exhaustion.
- Hierarchical skill system: Provides reusable strategy and access skills that enhance the search process and avoid repeating failed search patterns across runs.
Results
On the WideSearch and GISA benchmarks, SearchOS leads single-agent and multi-agent baselines across all evaluation metrics, paving the way for robust information-seeking collaboration.
--- *Auto-collected on 2026-07-18*