Toward the Science of Multi-Agent Communication: A Comprehensive Survey

Source abstract

Multi-agent systems now span autonomous driving, robotic teams, and collaborative LLM assistants, where inter-agent communication determines system-level performance. However, research on multi-agent communication has fragmented into paradigms that evolved in isolation, from Multi-Agent Reinforcement Learning (MARL) through Emergent Language (EL) to Large Language Model (LLM)-based agents, traditionally relying on continuous channels, discrete symbols, and natural language respectively. Existing surveys either focus on one paradigm or only compare paradigms along descriptive axes, overlooking the informational properties a message shares across paradigms. We present a unified survey of multi-agent communication spanning MARL, EL, and LLM-based agents. We first model communication as a state-perturbation operator on the receiver’s state ⟨h, θ⟩, giving cross-paradigm analysis a common mathematical ground. From this operator we derive Format, pairing it with an independent Content axis to form a unified Format Content taxonomy. Format is the architectural depth at which a signal couples to the receiver, spanning raw sensory signals, discrete symbols, latent vectors, and parameter updates; Content is the semantic payload, partitioned by referent, internal or external, and function, descriptive or prescriptive. We further propose the Communication Pentagram, which shifts evaluation from the downstream task to the channel itself along five intrinsic, task-independent dimensions: Bandwidth, Computational Cost, Robustness, Generalization, and Interpretability. Finally, we derive cross-paradigm design principles: a message’s Format-Content pair constrains the operational choices. Together, these reveal shared design patterns across historically separate paradigms, enabling cross-paradigm comparison and a predictive roadmap for next-generation agent systems.