Skip to content
WeChat · Weijiqi Lixin·Indexed by this site · Originally published· AI updates →· Its place in the scan record →

National University of Singapore proposes EvolveMAS: without changing model parameters, enabling multi-agents to continuously evolve in collaboration, negotiation, and adversarial scenarios

新加坡国立大学:不用更新大模型参数,多智能体也能持续进化

AI summary

A team from the National University of Singapore proposes in an arXiv paper a multi-agent self-evolution framework, EvolveMAS, which does not fine-tune model parameters and relies only on test-time memory governance, adapting simultaneously to three types of scenarios—collaboration, negotiation and adversarial—and achieving a normalized average score of 77.2 on 5 benchmarks, 9.7 points higher than the strongest scenario-specific baseline. The framework binds each main agent one-to-one with an independent MemoryAgent; three layers of structured memory—episodic, semantic and working—are isolated outside the main agent's context, and it evolves continuously on two time scales: single-episode online adaptation and cross-episode knowledge evolution. Ablation experiments show that on the LLM-Deliberation negotiation task, the independent MemoryAgent raises the consensus rate from 24% to 74%, reaching 86% when cross-episode knowledge is added; EvolveMAS's context usage stays at 0.62%, below the 6.54% of ordinary schemes. The paper also notes limitations: in zero-sum adversarial settings, both sides evolving simultaneously cancels out each other's gains, and weak base models struggle to abstract general strategies. The project is open-sourced at https://github.com/WuYalun/EvolveMAS

Read originalmp.weixin.qq.com