Multi-agent debate topologies — MultiAgentBench (ACL 2025)
star
hub
fast-factual wins here
chain
stepwise refinement
tree
hierarchical, divide-and-conquer
graph
research wins here
measured results from MultiAgentBench (MARBLE, ACL 2025)
graph best for research; +3% milestone achievement over next-best topology
coordination tax: past ~4 agents, wall-clock and token cost grow faster than accuracy
heterogeneity (different base models) consistently beats adding more homogeneous agents
bounded rounds (2-3) matter: unbounded debate rewards conformity, not correctness
star is the cost sweet spot when the hub is high-capability and workers are cheap