MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems
MANTA lets a multi-agent system change its own communication structure while it is running.
The framework starts with a task-conditioned topology, then monitors collaboration traces during inference. When the current setup is not enough, it can make bounded changes to roles, links, execution order, information visibility, and validation paths without changing the task interface or agent budget. The authors report testing it on five benchmarks, where it reached a 74.0 average score and beat the strongest baseline by 5.8 points. ArXiv · AI/CL/LG's note
The framework starts with a task-conditioned topology, then monitors collaboration traces during inference. When the current setup is not enough, it can make bounded changes to roles, links, execution order, information visibility, and validation paths without changing the task interface or agent budget. The authors report testing it on five benchmarks, where it reached a 74.0 average score and beat the strongest baseline by 5.8 points. ArXiv · AI/CL/LG's note
score 5