Active Learning for Communication Structure Optimization in LLM-Based Multi-Agent Systems

Researchers propose an active learning framework for optimizing communication structures in multi-agent systems powered by large language models, using ensemble-based task selection to identify the most informative training tasks while reducing token consumption and computational costs.

This research addresses a critical inefficiency in LLM-based multi-agent systems by tackling how agents communicate with each other. Rather than treating all training tasks equally, the proposed method intelligently selects which tasks provide the most value for teaching agents how to optimize their communication patterns. This matters because multi-agent LLM systems face escalating token costs and performance variability—two major barriers to practical deployment in production environments.

The technical approach leverages ensemble Kalman inversion to estimate task informativeness without requiring derivatives, making it particularly suited for black-box systems where agent behavior remains opaque. By treating communication structure optimization as a Bayesian parameter estimation problem, researchers can measure how much each task would shift the distribution over possible agent interaction patterns. The framework then uses embedding-based representative selection to create a manageable candidate pool and combines this with surrogate modeling and Thompson sampling for scalability.

The practical implications extend across industries deploying multi-agent AI systems. Organizations running language model-based agent teams can potentially reduce infrastructure costs substantially by training more efficiently under limited budgets. The method's robustness in adversarial settings—including scenarios with agent attacks—suggests applicability to security-critical deployments. This research represents incremental but meaningful progress in making multi-agent LLM systems more economically viable and stable, though it remains a technical contribution without direct market-moving implications.

• →Active learning framework reduces computational budget requirements for optimizing multi-agent LLM communication structures.
• →Ensemble Kalman inversion enables task informativeness estimation without derivatives, suitable for black-box systems.
• →Method demonstrates effectiveness even under adversarial conditions with agent attacks.
• →Embedding-based candidate pool selection enhances scalability of the optimization process.
• →Approach simultaneously improves downstream performance while decreasing token usage.