Self-Evolving Multi-Agent Framework for Efficient Decision Making in Real-Time Strategy Scenarios

A research team led by Li Ma, Hao Peng, and Yiming Wang has introduced SEMA (Self-Evolving Multi-Agent), a novel framework designed to solve the speed-quality dilemma for AI in Real-Time Strategy (RTS) games like StarCraft II. Traditional Large Language Models (LLMs) struggle here, as expansive game states cause prohibitive inference delays, and stochastic planning errors undermine logical consistency. SEMA tackles this with a collaborative multi-agent system that facilitates self-evolution, adaptively calibrating model bias through in-episode assessment and cross-episode analysis.

A key innovation is dynamic observation pruning based on structural entropy, which models the game state topologically. This distills high-dimensional data into core semantic information, drastically cutting down the processing load. The framework also employs a hybrid knowledge-memory mechanism, integrating micro-trajectories, macro-experience, and hierarchical domain knowledge to enhance both strategic adaptability and decision consistency. In experiments across multiple StarCraft II maps, SEMA validated its efficiency, achieving superior win rates while reducing average decision latency by over 50% compared to previous approaches.

This work represents a significant step beyond using LLMs as simple planners. By creating a specialized, evolving multi-agent architecture, the researchers have built a system that can reason and act under the severe time pressure of an RTS environment. The success in StarCraft II, a benchmark for AI complexity, demonstrates a path forward for deploying advanced AI in other dynamic, real-time domains where speed and strategic depth are equally critical.