DiffMAS: Training Framework for Latent Communication in Multi-Agent LLM Systems

A new training framework called DiffMAS has been introduced by researchers, which integrates latent communication as a learnable aspect in multi-agent systems utilizing large language models. In contrast to traditional methods that depend on static text-based protocols, DiffMAS leverages internal representations like key-value caches for communication between agents. This framework enables parameter-efficient supervised training across multi-agent latent trajectories, facilitating the joint learning of information encoding and interpretation during interactions. Experiments conducted on mathematical reasoning, scientific question answering, code generation, and commonsense benchmarks demonstrate notable enhancements in reasoning accuracy and decoding. This study fills a void in current multi-agent LLM research, which mainly concentrates on agent roles and orchestration while treating communication as a fixed interface.