Renmin, Ant Debut SearchOS-V1 for Grounded Web Research

Most of the multi-agent search frameworks that came out this year leaned harder on more agents or bigger context windows. SearchOS-V1, described in a new paper from Renmin University's Gaoling School of Artificial Intelligence and Ant Group, is one of the few that instead tries to fix the state-tracking problem: once a research agent has issued dozens of searches, it forgets what it already covered, which dead ends it hit, and which values it actually grounded to which sources.

The formulation the authors reach for is old-database-textbook stuff. They treat an open-domain question as "relational schema completion with grounded citations," where each answer cell holds a value plus a source URL and an anchored excerpt. Around that they add a shared state layer called SOCM that holds a Frontier Task pool, an Evidence Graph, a Coverage Map, and a Failure Memory, updated through locked read-modify-writes so agents never observe stale progress. A middleware harness sits between the agent and its tools, intercepting model and tool calls to inject state, extract evidence, and detect stalls or budget overruns without asking the agent to remember to check.

The reported numbers on the two benchmarks they test are the reason to pay attention. On WideSearch, a 200-question benchmark drawn from more than 15 domains, SearchOS attains 80.3 item-level F1 against 76.0 for the strongest baseline, A-MapReduce. On GISA, a 373-query set, it leads across every question type and hits 76.5 F1 on Set questions versus 63.1 for the best baseline, a 13.4-point jump. A scheduling ablation, running with an 8-slot concurrency cap, shows continuous dispatch cutting end-to-end time by 24.3% over synchronized batch scheduling while also raising Item F1.

The honest caveat is that the backbone is GLM-5 with Qwen3.5-35B-A3B doing evidence extraction, and results are reported as best-of-three (Max@3), which softens variance. The system also ships with 280 hand-authored skills spanning strategy and site-specific access, and the paper does not isolate how much of the win comes from that library versus SOCM itself. Per-query cost, code availability, and behavior on domains outside the seeded skill catalog are not addressed in the sections I read.

For anyone building a research assistant on frontier LLMs, the piece worth stealing is not the framework name. It is the argument that search progress belongs in a system-managed store with verifiable provenance, not in the model's rolling context. That is the direction the interesting agent work is going.