Xiamen University's FlowMAS uses GFlowNet to auto-design multi-agent topologies
jiqizhixin · x · 2026-10-24
Xiamen University's ASC Lab presents FlowMAS (accepted at NeurIPS 2026, arXiv 2609.37151), reframing multi-agent architecture design as a learnable generation problem.
- Pain point: forming an agent team is easy, but designing the topology — who talks to whom, in what order — is hard; hand-designed workflows are rigid and rarely optimal
- Method: Generative Flow Networks (GFlowNet) as the engine to discover diverse, high-performing workflow topologies, paired with curiosity-driven exploration and information-theoretic guidance
- Operators include CoT, testing, vote-ensemble, decomposition, aggregation, with chain/fan-in/fan-out DAG primitives
- First approach making AI team construction simultaneously automatic, diverse, and high-performing
More from coding & agent
- Lee Rob's 45-minute masterclass explains how always-on AI agents work — Saboo_Shubham_ · 2026-10-24
- Prompt engineering is 'dead', long live prompt engineering: a cyclist-recipe prompt shows the craft persists — chrisalbon · 2026-10-24
- Few-shot prompting trick: turn examples into a list the system completes — mmitchell_ai · 2026-10-24
- Rewriting the world in Rust: nearly 1T tokens and 50k+ commits per month — doodlestein · 2026-10-24
- NeurIPS paper OAT: attribute agent failures by learning only from success flows — SharonYixuanLi · 2026-10-24
- Claude writes its own parallel-agent prompt and nails the brief, screenshot goes viral — hewarsaber · 2026-10-24