Reproduce “Self-Evolving Multi-Agent Framework for Efficient Decision Making in Real-Time Strategy Scenarios” on OpenResearch

About this paper

Li Ma, Hao Peng, Yiming Wang, Hongbin Luo, Jie Liu, Kongjing Gu, Guanlin Wu, Hui Lin and 1 more

Researchers from Beihang University and collaborators introduce SEMA, a self-evolving multi-agent framework that enables efficient and robust decision-making for Large Language Models (LLMs) in Real-Time Strategy (RTS) scenarios. SEMA significantly reduces LLM inference latency by over 50% and input token load by more than 70% while achieving high win rates, including 100% on StarCraft II Melee maps against built-in AI.

Key insights

• Dynamic observation pruning based on structural entropy can drastically reduce LLM input token load (over 70%) and inference latency (over 50%) in complex environments without sacrificing critical strategic information.
• A multi-agent collaborative framework combined with a hybrid knowledge-memory system effectively mitigates LLM stochasticity, enhancing decision consistency and strategic adaptability through real-time experience retrieval and contextual input.
• A self-evolving mechanism, driven by post-match reflection and iterative experience updates, enables LLM-based agents to continuously refine strategies and improve performance in challenging, adversarial RTS scenarios.

Summary by alphaXiv.