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

2026-03-26 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Gaming & Interactive Media · Depth: Expert, long

Summary

SEMA (Self-Evolving Multi-Agent) is a novel framework designed to enhance decision-making in Real-Time Strategy (RTS) scenarios like StarCraft II, addressing the speed-quality trade-off inherent in Large Language Models (LLMs). The framework employs a collaborative multi-agent system that self-evolves by adaptively calibrating model bias through in-episode assessment and cross-episode analysis. It integrates dynamic observation pruning based on structural entropy to topologically model game states, distilling high-dimensional data into core semantic information and significantly reducing inference time. Additionally, SEMA utilizes a hybrid knowledge-memory mechanism combining micro-trajectories, macro-experience, and hierarchical domain knowledge to improve strategic adaptability and decision consistency. Experiments on multiple StarCraft II maps show SEMA achieves superior win rates and reduces average decision latency by over 50%, demonstrating its efficiency and robustness.

Key takeaway

For research scientists developing autonomous agents in dynamic, high-dimensional environments like RTS games, SEMA offers a blueprint for overcoming LLM latency and inconsistency. You should consider integrating structural entropy-driven observation pruning and a multi-agent collaborative architecture to achieve real-time, high-quality decision-making without extensive fine-tuning. This approach can significantly improve agent robustness and performance in complex adversarial settings.

Key insights

SEMA is a multi-agent LLM framework for RTS that uses structural entropy and hybrid memory to achieve fast, consistent decisions.

Principles

• Adaptive bias calibration improves LLM consistency.
• Topological modeling reduces high-dimensional observation data.
• Hybrid memory enhances strategic adaptability.

Method

SEMA employs a decision-making, evaluation, and policy agent system. It uses structural entropy for dynamic observation pruning, and a hybrid knowledge-memory mechanism for cross-round strategy evolution and real-time action correction.

In practice

• Apply structural entropy for data compression in dynamic environments.
• Implement multi-agent systems for complex task decomposition.
• Utilize hybrid memory for long-term strategic consistency.

Topics

• Real-Time Strategy
• Multi-Agent Systems
• Large Language Models
• Structural Entropy
• StarCraft II

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• SOM: Structured Opponent Modeling for LLM-based Agents via Structural Causal Model — SOM improves LLM agent prediction by separating opponent modeling and prediction using a Structural Causal Model.
• AEM: Adaptive Entropy Modulation for Multi-Turn Agentic Reinforcement Learning — AEM uses response-level entropy to adaptively balance exploration and exploitation in multi-turn LLM agent RL.
• Multi-agent Framework for Time-Sensitive Complementary Collaboration in Minecraft — Large language models struggle with real-time, dynamic multi-agent collaboration due to latency and coordination complexity.
• SEMA-RAG: A Self-Evolving Multi-Agent Retrieval-Augmented Generation Framework for Medical Reasoning — Multi-agent RAG with self-evolving, sufficiency-driven retrieval significantly improves medical question answering accuracy.
• NeuroMAS: Multi-Agent Systems as Neural Networks with Joint Reinforcement Learning — NeuroMAS treats multi-agent systems as trainable neural architectures, learning communication and specialization via reinforcement learning.
• Enhancing Decision-Making with Large Language Models through Multi-Agent Fictitious Play — Iterative fictitious play with LLM agents resolves "stance entanglement" in complex decision-making by decomposing recursive anticipation into sequential best responses.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.