Latent Collaboration in Multi-Agent Systems (github.com)

LatentMAS introduces a groundbreaking multi-agent reasoning framework that shifts the communication among agents from token space into the model's latent space, enabling efficient collaboration through the exchange of latent thoughts instead of lengthy textual traces. This innovative approach significantly reduces token consumption by 50-80% and achieves impressive wall-clock speed improvements, reaching speeds three to seven times faster than traditional text-based multi-agent systems. The framework is also compatible with any Hugging Face model and supports vLLM backends, making it broadly applicable across varied architectures. The significance of LatentMAS for the AI/ML community lies in its potential to streamline multi-agent interactions, facilitating faster reasoning and decision-making processes in complex systems. Additionally, new community-driven extensions, such as Science-LatentMAS for scientific applications and kNN-latentMAS for memory efficiency, enhance its capabilities, promoting a modular and flexible environment for diverse agent types. As a spotlight paper accepted at ICML 2026, LatentMAS is poised to influence future research directions in multi-agent systems and collaborative AI, ultimately paving the way for more advanced and efficient AI-driven tasks across multiple domains.