MAS-Loop: Beyond One-Shot Orchestration towards Self-Evolving Multi-Agent Systems

DOI10.48550/ARXIV.2610.38946MaRDI QIDQ7414973arXivFDO

Authors Chengwei Qin, Hehai Lin, Qi Yang

Publication date 28 October 2026

Published in arXiv

Tldr This paper closes the feedback loop in training-time multi-agent system (MAS) design by coupling an Orchestrator with a Refiner to execute and update MAS iteratively based on trajectories rather than only terminal outcomes; experiments show significant performance gains from this approach across various benchmarks.

Full work available at URL https://doi.org/10.48550/ARXIV.2610.38946

arXiv classification

cs.AIcs.MA

Cites work

• Degrees of Freedom in Multi-Agent Coordination: Between Centralized and Decentralized Regimes

This page was built for publication: MAS-Loop: Beyond One-Shot Orchestration towards Self-Evolving Multi-Agent Systems

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7414973)