Latent Collaboration in Multi-Agent Systems

Multi-agent systems (MAS) extend large language models (LLMs) from independent single-model reasoning to coordinative system-level intelligence. While existing LLM agents depend on text-based mediation for reasoning and communication, we take a step forward by enabling models to collaborate directly within the continuous latent space. We introduce LatentMAS, an end-to-end training-free framework that enables pure latent collaboration among LLM agents. In LatentMAS, each agent first performs auto-regressive latent thoughts generation through last-layer hidden embeddings instead of text. Then, a shared latent working memory preserves and transfers each agent's internal representations and latent thoughts, ensuring lossless information exchange without re-encoding. We provide detailed theoretical analyses showing that LatentMAS achieves higher expressiveness and lossless information preservation with lower overall complexity than standard text-based MAS. In addition, empirical evaluations across 9 comprehensive benchmarks spanning math and science reasoning, commonsense understanding, and code generation show that LatentMAS outperforms advanced single agents and text-based MAS baselines, achieving up to 14.6\% higher accuracy, reducing output token usage by 70.8\%-83.7\%, and providing 4$\times$-4.3$\times$ faster end-to-end inference.

Large language models are increasingly used as teams of specialized agents, where one model may plan, another may critique, and another may solve the task. However, most multi-agent systems make these agents communicate by writing text to each other, which can be slow, expensive, and may compress away useful internal information. We introduce LatentMAS, a framework that lets language-model agents collaborate directly through their internal hidden representations instead of repeatedly converting their thoughts into text. Each agent performs reasoning in this continuous “latent” space, and then passes its internal working memory to the next agent without requiring additional training. Across math, science, commonsense, and coding tasks, LatentMAS improves reasoning accuracy while using far fewer generated tokens and running substantially faster than text-based multi-agent systems. This suggests a more efficient way to build collaborative AI systems that can reason together without relying entirely on natural language communication.