EvoMAS : Learning Execution-Time Workflows for Multi-Agent Systems
Chengdong Xu, Kaiqiang Ke, Ziheng Liu, Jiaqi Wei, Zibo Shao, Weile Guo, Chao Yu, [0.3em] Sun Yat-Sen University · 2026
Read EvoMAS : Learning Execution-Time Workflows for Multi-Agent Systems by Chengdong Xu, Kaiqiang Ke, Ziheng Liu, Jiaqi Wei, Zibo Shao, Weile Guo, Chao Yu, [0.3em] Sun Yat-Sen University (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-08769.txt.
Description
Chengdong Xu and six coauthors build multi agent workflows while a task runs, rather than freezing the design before execution starts. Long tasks shift subgoals and information needs stage by stage, so one shot workflow design fits poorly. EvoMAS treats workflow building as a sequence of decisions along one trajectory, constructing an explicit task state at each stage through planning, evaluation, and updating steps.
A learned adapter then assembles a stage specific layered workflow from a fixed agent pool, trained mainly on final verifiable success. Across research and question answering benchmarks, EvoMAS beats single agents and recent automatic design methods. Coordination that adapts as evidence arrives proves stronger than coordination fixed in advance.
Catalog record
| CREATOR | Chengdong Xu, Kaiqiang Ke, Ziheng Liu, Jiaqi Wei, Zibo Shao, Weile Guo, Chao Yu, [0.3em] Sun Yat-Sen University |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08769