Xiamen team's FlowMAS learns agent workflow topologies with GFlowNets, beats AFlow at 15x less compute

机器之心 · wechat · 2026-10-16

A team from Xiamen University's ASC Lab landed a NeurIPS 2026 paper, FlowMAS, that turns multi-agent workflow topology design into a learnable generative problem.

Key ideas

Uses GFlowNets (originally from Bengio's group for molecule generation) as the generative backbone: sampling probability proportional to each architecture's reward, preserving diversity and fixing the homogeneous outputs of VAE/diffusion-based designers.
Adds intermediate rewards at every construction step, rederiving flow-conservation and trajectory-balance constraints.
Two scoring modules: curiosity (random network distillation) rewards novel structures; an information-guided module uses partial information decomposition to reward synergistic contributions plus Gram-matrix spectral entropy to keep all communication channels useful. Scoring is training-only, zero inference cost.

Results

First place on 6 benchmarks (GSM8K/MATH, HumanEval/MBPP, HotpotQA/DROP) across GPT-4o-mini, Llama-3.1-8B, Gemma-4-26B vs 15 baselines; 70.50 avg on Llama-3.1-8B vs MaAS 67.73 and AFlow 67.55.
+8.17 over hand-designed and +6.12 over automated methods on average; biggest gains on the 8B model.
Cost: AFlow uses 15x+ inference tokens with 4.5 points lower accuracy; ADAS needs 220x training tokens.
Ablations: swapping the backbone or removing either module drops HumanEval from 69.6 to 62.8.

Paper: https://arxiv.org/abs/2609.37151

Original post →

Xiamen team's FlowMAS learns agent workflow topologies with GFlowNets, beats AFlow at 15x less compute

More from coding & agent