LatentMAS
LatentMAS lets multiple AI agents collaborate by passing latent representations through a shared working memory instead of exchanging text, aiming to reduce token usage and speed up multi-step reasoning.
LatentMAS is a multi-agent reasoning framework that moves collaboration between agents out of token space and into a model's latent space. Instead of writing out long text reasoning traces, agents pass latent thoughts through their own working memory to communicate with each other.
The approach is training-free and works with any Hugging Face model, with an optional vLLM backend for serving. The project reports that this design lowers token usage and produces faster wall-clock reasoning times compared with text-based multi-agent exchange, while aiming for comparable or better task performance.