TacoMAS: Test-Time Co-Evolution of Topology and Capability in LLM-based Multi-Agent Systems

2026-05-10

AI Summary

Existing large language model (LLM)-driven multi-agent systems struggle to jointly optimize agent capabilities and communication topology during inference, limiting their adaptability to diverse tasks. This work proposes TacoMAS, a novel framework that, for the first time, enables dual-scale co-evolution of both capabilities and topology at inference time: a fast loop dynamically refines agent specializations based on trajectory feedback, while a slow loop, orchestrated by a meta-LLM, performs structural edits such as agent addition/removal and connection rewiring. Theoretically grounded in convergence to a task-specific stable equilibrium, TacoMAS integrates online graph adaptation, trajectory-level feedback learning, and dynamic topology optimization. Empirical results demonstrate that TacoMAS outperforms the strongest baseline by an average of 13.3% across four benchmarks, significantly surpassing nearly 20 existing methods.

Abstract

Multi-agent systems (MAS) have emerged as a promising paradigm for solving complex tasks. Recent work has explored self-evolving MAS that automatically optimize agent capabilities or communication topologies. However, existing methods either learn a topology that remains fixed at inference time or adapt only the topology or capability during inference. We empirically and theoretically show that effective test-time evolution requires jointly adapting both axes, but on different time scales: capabilities should update rapidly to handle emerging subtasks, while the topology should evolve more slowly to preserve coordination stability. We then introduce TacoMAS, a test-time co-evolution framework for dynamic MAS. TacoMAS formulates MAS inference as a task of online graph adaptation, where nodes represent agents with role-specific capabilities and edges define their communication topology. During inference, a fast capability loop updates agent expertise using trajectory-level feedback, while a slow meta-LLM-driven topology loop performs agents' birth-death operations on MAS, including edge edit, agent addition, and agent removal. We further show that this fast-slow design drives MAS evolution toward a task-conditioned stable equilibrium. Experiments on four benchmarks demonstrate that TacoMAS outperforms nearly 20 multi-agent baselines, achieving an average improvement of 13.3% over the strongest baseline. The codes are released at https://github.com/chenxu2-gif/TacoMAS-MultiAgent.

Problem

• multi-agent systems
• test-time adaptation
• topology evolution
• capability adaptation
• LLM-based agents

Innovation

• test-time co-evolution
• multi-agent systems
• capability-topology adaptation
• fast-slow dynamics
• online graph adaptation