StateBridge: Training-free Hidden-state Alignment for Latent Communication in LLM Multi-Agent Systems

StateBridge enables latent communication between LLM agents by aligning a sender's final-layer hidden states to a receiver's input space using a closed-form orthogonal transformation. This training-free method uses norm calibration and vocabulary anchoring to maintain compatibility with pretrained distributions without requiring new projectors or layer-by-layer memory injection.