LatentMAS – agent collaboration from token space into the model's latent space (arxiv.org)

Researchers have introduced LatentMAS, a pioneering framework that enhances multi-agent collaboration among large language models (LLMs) by enabling interactions directly within a model's latent space. Unlike traditional approaches that rely on text-based communication, LatentMAS facilitates pure latent collaboration, where agents utilize hidden embeddings to generate auto-regressive thoughts and share internal representations through a shared latent working memory. This cognitive mechanism ensures lossless information exchange, significantly improving system-level intelligence. The significance of LatentMAS lies in its ability to achieve higher expressiveness and efficiency compared to conventional multi-agent systems. The framework demonstrates impressive empirical results, outperforming existing single-model and text-based multi-agent baselines across nine benchmarks related to reasoning, commonsense understanding, and code generation. Notably, it yields up to 14.6% higher accuracy while drastically reducing output token usage by 70.8%-83.7% and enhancing inference speeds by 4 to 4.3 times. The open-sourced nature of the code and data further encourages community exploration and development, marking a substantial advancement in the efficiency and effectiveness of collaborative AI systems.