Active Learning Framework Optimizes Communication in LLM-Based Multi-Agent Systems

Researchers have introduced a novel active learning framework designed to optimize the communication structure of Large Language Model-based Multi-Agent Systems (LLM-MAS). While optimizing these structures improves performance and reduces token usage, existing methods relying on random task sampling often suffer from instability and sensitivity to training sets due to varying task difficulty. To address this, the proposed method employs an ensemble-based information-theoretic task selection framework. It estimates task informativeness by measuring how candidate tasks alter the distribution over graph parameters, utilizing ensemble Kalman inversion as an efficient, derivative-free approximation for Bayesian updates. This approach is particularly effective for black-box and noisy systems. To ensure scalability, the framework constructs a compact candidate pool via embedding-based representative selection and integrates surrogate modeling with batch Thompson sampling. Validated in both benign environments and scenarios involving agent attacks, the method demonstrates significant effectiveness in optimizing communication structures under constrained computational budgets, offering a robust solution for enhancing the efficiency and stability of complex multi-agent interactions.