StateBridge passes hidden states between agents with a closed-form orthogonal map — no training, no projector

arXiv·medium signalXBlueskyLinkedInCopy link

StateBridge (arXiv 2608.13317, 2026-08-13) attacks the discrete bottleneck in multi-agent systems: converting a sender's continuous hidden states into text tokens throws away information token identities cannot carry. Prior latent-communication work either injects working memory layer by layer through the transformer or requires trained projectors that don't port across models; StateBridge instead aligns the sender's final-layer hidden states to the receiver's input space with a closed-form orthogonal transformation, adds lightweight norm calibration and vocabulary anchoring, and prepends the aligned states as a continuous prefix. Across math reasoning, code generation and QA with four models from two families it takes best or tied-best on 22 of 26 model-task pairs.