StateBridge: Training-free Latent Communication for LLM Multi-Agent Systems

A recent paper published on arXiv (2608.13317) presents StateBridge, a novel approach for latent communication in multi-agent systems utilizing large language models (LLMs) that does not require training. Conventional multi-agent systems rely on discrete text tokens, which can limit information flow by overlooking details from continuous hidden states. While latent communication allows agents to share hidden representations directly, current methods either necessitate trained projectors or involve adding working memory layer by layer across transformers, hindering portability. StateBridge utilizes a closed-form orthogonal transformation to align the sender's final-layer hidden states with the receiver's input space, incorporating lightweight norm calibration and vocabulary anchoring for compatibility with the pretrained input distribution. This alignment is prepended to the receiver's input, facilitating integration without further training and enhancing communication efficiency and richness among LLM agents, which may lead to improved performance in multi-agent tasks.