Latent communication speeds up multi agent system interactions
Thought Blocks: Non-Autoregressive Latent Communication for Multi-Agent Systems via Trajectory-to-State Compilation
Multiagent Systems
Summary
Multi-agent systems, where many AI agents talk to each other, often send messages as text, which can be slow. The authors show that these agents can share information faster by creating special blocks of memory called Thought Blocks, which capture what they want to say all at once instead of step by step. This approach greatly reduces the time and amount of data needed during communication. Their method speeds up interactions up to 8.6 times compared to text and avoids slow sequential message building.
What this means in practice
- •For ai system engineers: Implement faster communication among AI agents in systems by directly generating latent state messages in one step.
- •For robotics developers: Reduce latency in robot team coordination by using non-sequential latent communication for state sharing.
Authors
Zhenhan Guan, Xiaolong Liu, Hengwei Ye, Zheng Tian
Abstract
LLM-based multi-agent systems typically communicate through text, incurring repeated generation and re-encoding costs. Although latent communication removes explicit verbalization, constructing latent messages may still require autoregressive latent trajectories. We therefore distinguish message representation from message construction and ask whether agents can directly construct useful continuation states without executing the trajectories that normally produce them. We formulate this problem as trajectory-to-state compilation and introduce Thought Blocks, fixed-size blocks of latent key--value (KV) continuation memory constructed in one shot. Thought Blocks achieve a 2.5 times average latency speedup over autoregressive latent communication. Compared with text-based multi-agent communication, Thought Blocks reduce text tokens by 45.7% and autoregressive steps by 84.8%, achieving an 8.6 times average latency speedup. These results demonstrate that trajectory-to-state compilation enables efficient inter-agent latent communication without sequential latent rollouts at inference.