Multi-agent system improves long-step evidence search and answer quality
Harness-Search: Guiding Long-Horizon Search through Multi-Agent Coordination
Computation and Language
Summary
When machines need to find answers by gathering information step-by-step over a long time, they can get confused or stop too early. The authors created a system that splits this search into three teamwork roles: one finds information, another checks and saves it, and the last decides if enough has been found. This teamwork helps keep the search on track and collect better evidence. Tests show their system finds more useful information and makes better final answers than older methods.
What this means in practice
- •For knowledge graph engineers: Build systems that gather and verify evidence across multiple retrieval steps to improve the accuracy of automated knowledge graph population.
- •For customer support ai developers: Develop AI that better coordinates search and validation steps to provide more reliable and well-supported answers from large document sets over long queries.
Authors
Shanyong Wang, Zhenwen Ji, Lei Jin, Yining Zhao, Yicheng Qian, Chengqiang Lu, Yi Wu, Yao Hu, Lizhen Cui, Yanyu Xu
Abstract
Long-horizon search requires agents to gather evidence across multiple steps and synthesize it into well-supported answers. The recent agent harnesses provide a natural and promising framework to support such long-running search processes. As interaction histories grow, one single agent in harnesses might get stuck and cause the policy to lose track of unresolved questions, overlook useful evidence, or terminate before sufficient support has been collected. One of promising way is to decouple three distinct responsibilities of proposing retrieval actions, updating persistent state, and deciding when to stop rather than concentrating them within a single policy. Targeted at it, we introduce Harness-Search, a multi-agent search harness to reduce the local errors propagating across subsequent exploration, evidence curation, and termination decisions. In particular, Harness-Search assigns these responsibilities to three permission-bounded authorities: a Retrieval Policy that proposes search operations, a Memory Operator that validates and commits persistent-state updates, and a Summary Auditor that accepts or rejects termination based on the sufficiency of the curated evidence. Together, these roles form a Propose-Commit-Audit loop in which actions are proposed, persistent evidence is selectively committed, and stopping decisions are subjected to an explicit sufficiency check. Across seven long-horizon search benchmarks, Harness-Search improves both retrieval and answer generation under the same policy backbone, increasing Recall by 4.60-27.92 points and Final-Answer Recall by 12.34-30.13 points over the strongest harness-based baseline on each evidence-retrieval benchmark. Moreover, trajectory-level analyses show that Harness-Search continues to accumulate useful evidence and expand evidence coverage with less redundant retrieval as the search history grows.