DiffMAS: Differentiable End-to-End Optimization of Multi-Agent Communication via Latent KV-Cache Sharing

arXiv·medium signal

Researchers propose DiffMAS, a training framework for jointly optimizing multi-agent communication with reasoning by using latent key-value cache representations instead of text-based inter-agent messaging. Most multi-agent LLM systems treat communication as a fixed text interface; DiffMAS demonstrates that latent communication through internal representations can be jointly optimized with reasoning, potentially enabling more efficient multi-agent coordination than prompt-engineering approaches.