LatentMAS

LatentMAS is a multi-agent reasoning framework that shifts agent collaboration from the text space to the model's latent space. Agents communicate by passing latent thoughts instead of generating lengthy textual reasoning traces. The framework offers high efficiency, training-free latent alignment, and generalizability, compatible with any Hugging Face model and optionally supporting the vLLM backend. LatentMAS achieves outstanding performance in multi-agent systems with reduced token usage and significant speed improvements.