Unifying information-theoretic and learning-based approaches for goal-oriented MAS communication

Establish a hybrid framework that bridges information-theoretic formulations (e.g., mutual information, rate–distortion, semantic rate–distortion, information bottleneck) with learning-based communication strategies (e.g., reinforcement learning, attention-based message passing) for goal-oriented communication in multi-agent systems.

Background

Information-theoretic models provide principled tools for quantifying relevance and compression but typically assume known models and distributions, whereas learning-based approaches discover effective communication policies empirically but lack formal guarantees.

The paper explicitly identifies the need to close this gap, motivating the development of hybrid methods that integrate theoretical structure with data-driven learning for task-aligned, efficient multi-agent communication.