Hehai Lin and coauthors propose MAS-Loop for revising multi-agent systems

Hehai Lin and coauthors propose MAS-Loop for revising multi-agent systems

Hehai Lin, Qi Yang, and Chengwei Qin argue that many training-time multi-agent system (MAS) approaches are one-shot: an Orchestrator learns from a system’s terminal outcome but does not use its execution trajectory to revise the system. Their proposed MAS-Loop pairs an Orchestrator, which builds and runs a system, with a trajectory-conditioned Refiner. The Refiner uses the trajectory to localize an actionable failure, diagnose it, and propose a bounded patch to the agents, information flow, or execution configuration. The updated MAS is then re-executed on the same task, allowing the authors to compare its outcome with the source MAS under stochastic execution. They say this comparison gives the Orchestrator and Refiner complementary learning signals, driving policy-level co-evolution across training iterations. The authors report strongest overall performance among evaluated baselines and consistent gains from refinement across all evaluated benchmarks. They also say the process yields MAS-JudgeCorpus, a dataset of execution-grounded source-updated pairs with orchestration- and trajectory-level supervision. The approach makes intermediate execution evidence a basis for revising systems and learning from observed weaknesses, rather than relying only on terminal outcomes. The abstract does not provide benchmark-specific results or the size and conditions of the reported gains.

Publication: 94aeedac-5e6b-4777-89bd-84c0aabad3c2

attribution — pass: The paper is correctly attributed to Hehai Lin, Qi Yang and Chengwei Qin. The method, performance results and dataset are framed as the authors' claims or reports, not as established fact.

Authors:Hehai Lin, Qi Yang, Chengwei Qin

summary — uncertain: The summary faithfully tracks the abstract, and the title and question fit. The draft does not say in plain terms what a multi-agent system or an execution trajectory is. The Orchestrator and Refiner are briefly explained. Phrases such as policy-level co-evolution, stochastic execution, and orchestration- and trajectory-level supervision are also left unexplained. A non-specialist may not grasp what is being built or why trajectories matter. Smallest repair: add one plain-language clause defining MAS (several AI agents working together on a task) and trajectory (the record of intermediate steps), keeping the exact terms.

date — pass: The originalPublished value of 2026-10-28 matches the v1 submission date on arXiv. The feed date of 2026-10-31 is later and is correctly not used.

[Submitted on 28 Oct 2026]

support — pass: Each claim matches the abstract verbatim in substance. The sentence on significance restates the abstract's own contrast between terminal outcomes and trajectories. The draft does not overstate the results and correctly notes that the abstract gives no benchmark-specific figures. The question follows from that gap.

MAS-Loop achieves the strongest overall performance among the evaluated baselines, with consistent gains from refinement across all evaluated benchmarks.

media — pass: The selected media is the verified arXiv abstract page for this same paper. It is the most direct way to access the source.