Xiamen University's FlowMAS Teaches Teams of AI Agents to Assign Their Own Roles

The ASC Lab team at Xiamen University has published FlowMAS, a study of multi-agent systems that appears under the NeurIPS 2026 banner. The headline of the Jiqizhixin article frames the idea figuratively: AI teams should learn "排兵布阵" on their own, meaning to deploy their troops and arrange the order of action. All the available excerpt tells us is who the authors are and what the lab does. It contains no benchmarks, no model sizes and no comparisons with baseline methods.

Who is behind the work

The first author is Wang Haitao, a doctoral student admitted in 2025. Master's student Liang Chenjing worked on the project with him. The supervisor and corresponding author is Associate Professor Shen Xiji. The co-authors also include Zhang Haipeng, Hu Jiawei, Wang Sicheng, senior researcher Mei Songzhu, Professor Wen Chenglu and Professor Wang Cheng.

The Spatial Perception and Computing Lab lists three research areas: agents built on large models, reinforcement learning and multi-agent systems. FlowMAS sits at the intersection of all three.

What problem this line of research addresses

The title and subtitle suggest what the work is about, but that is an inference, not a summary of the paper. Classic LLM-based multi-agent setups are usually assembled by hand. An engineer decides how many agents are needed, which one plans, which one writes code, which one checks the result, and in what order they pass context to each other. Such a setup works well on one task but transfers poorly to others. When the task type changes, the topology has to be redesigned.

The phrase about independently "deploying forces" points to an attempt to remove that manual step: the system should learn to build a workflow (the name includes the word "flow") for a specific task. The lab's stated focus on reinforcement learning makes it likely that a reward for the final result serves as the signal for this tuning. How exactly FlowMAS works does not follow from the text we have.

What is known so far

• Name: FlowMAS.
• Venue: NeurIPS 2026, according to the publication's headline.
• Institution: ASC Lab, Xiamen University.
• Topic: self-learned organization of teams of AI agents.
• Metrics, datasets, baseline models, cost of running: not given in the source.
• Code and weights: not mentioned in the available excerpt.

What to check when reading the full paper

For now this is an announcement without data, and it is too early to draw conclusions about the method's usefulness. When the preprint appears, a few things are worth checking.

First, what it is compared against. A convincing comparison should include both hand-built multi-agent setups and a strong single agent with the same token budget. Multi-agent systems often win simply because they spend more compute.

Second, the cost of training and inference. If the system learns to build its workflow through RL, it matters how many runs that takes and how much the number of model calls per task goes up.

Third, transfer. Does a system that has learned to arrange agents on one type of task produce a scheme that works on others without additional training?

Fourth, whether the code is open. For practitioners building agent pipelines, reproducibility matters more than a flashy name.

For those running agent systems in production, FlowMAS is for now a reason to follow the preprint, not to change your architecture. We will return to this when numbers appear.