The Big Picture
Coordinating many smaller agents can match the search accuracy of a much larger single model while improving as you add more parallel agents—giving a cost-efficient way to handle broad information tasks.
ON THIS PAGE
The Evidence
A lead agent that divides work among many smaller subagents reaches similar performance to a huge single agent when searching across broad topics. A 4-billion-parameter version of the system hit 40.0% item F1 on the benchmark, matching a 671-billion-parameter single-agent baseline. Performance kept improving as more subagents worked in parallel, showing that 'width'—many workers—can substitute for 'depth'—one very large model—on broad information tasks. multi-agent system
Data Highlights
1WideSeek-R1-4B achieved 40.0% item F1 on the WideSearch benchmark.
2The 4-billion-parameter multi-agent system matched performance of a 671-billion-parameter single-agent model on the same benchmark.
3System was trained on a curated dataset of 20,000 broad information-seeking tasks and showed consistent gains as the number of parallel subagents increased.
What This Means
Engineers building search and research assistants who need to balance cost, speed, and coverage should consider this approach because smaller models can be run in parallel to approach the accuracy of huge models. Technical leaders deciding between buying or scaling big models versus building multi-agent pipelines can use these results to explore cheaper, scalable alternatives. Researchers studying how to organize AI workers can use this as evidence that coordination strategies matter as task breadth grows. planning pattern
Not sure where to start?Get personalized recommendations
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Results are reported on a curated benchmark and may not fully reflect messy, adversarial, or real-time web environments. The approach needs infrastructure for orchestration and tool integration; coordination overhead could reduce gains in some deployments. Metrics focused on item-level accuracy; latency, cost-per-query, and reliability under failure modes need separate evaluation before production rollout. State Inconsistency
Methodology & More
WideSeek-R1 reorganizes broad information-seeking from one deep reasoning agent into a lead-agent plus many subagents. The lead agent splits a broad task into smaller subtasks, each subagent runs in an isolated context with specialized tools, and results are combined back. Both the lead and subagents are trained together using a multi-agent reinforcement learning setup on a dataset of 20,000 curated tasks, so the system learns how to delegate and merge work effectively. The key result is that a 4-billion-parameter shared model, when used as a coordinated multi-agent system, achieved 40.0% item F1 on the WideSearch benchmark—comparable to a single-agent model with 671 billion parameters. Performance improved as the team of subagents grew, demonstrating that adding parallel workers (width) can substitute for a single very large model (depth) on broad tasks. Practical implications include potential cost savings and easier specialization (small agents with focused tools), but real-world adoption will need attention to orchestration cost, tool reliability, and trust signals for agent behavior and evaluation. Handoff Pattern Tool Use Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Affiliated with Tsinghua University (top institution) and several authors with modest h-indices — stronger institutional signal despite arXiv venue.