StateBridge: Training-free Hidden-state Alignment for Latent Communication in LLM Multi-Agent Systems
StateBridge is a training-free latent communication approach for LLM multi-agent systems that aligns the sender's final-layer hidden states to the receiver's input space via a closed-form orthogonal transformation, with lightweight norm calibration and vocabulary anchoring. It was evaluated on math reasoning, code generation, and question answering with four models from two families, achieving the best or tied-best score on 22 out of 26 model-task pairs.
StateBridge introduces a training-free method for latent communication in LLM multi-agent systems. Instead of converting hidden states to discrete tokens, it aligns the sender's final-layer hidden states to the receiver's input space using a closed-form orthogonal transformation, with norm calibration and vocabulary anchoring to maintain compatibility with the pretrained input distribution. The aligned states are prepended as a continuous prefix to the receiver's input. Evaluation on math reasoning, code generation, and question answering across four models from two families shows StateBridge achieves best or tied-best performance on 22 of 26 model-task pairs.
StateBridge eliminates the need for trained projectors by using a closed-form orthogonal transformation to align hidden states across models. The use of norm calibration and vocabulary anchoring suggests a focus on preserving the statistical properties of the pretrained input distribution, which may improve robustness and portability. The approach is evaluated across multiple model families, indicating potential for cross-model latent communication without fine-tuning.
Training-free latent communication could reduce the overhead of multi-agent LLM systems by avoiding tokenization bottlenecks and enabling more efficient information transfer. This may lower barriers to deploying heterogeneous multi-agent systems, as no additional training is required for alignment. The method's portability across model families could accelerate adoption in agentic workflows.
StateBridge could reduce inference costs and latency in multi-agent systems by avoiding token generation for inter-agent communication. It may enable more efficient collaboration between heterogeneous models, potentially lowering integration costs for enterprises building agent-based applications. The training-free nature removes the need for custom alignment training, simplifying deployment.
Next signals to watch include whether StateBridge is adopted in open-source multi-agent frameworks, whether follow-up work extends the approach to more than two agents or longer conversations, and whether the method scales to larger models or different modalities. Independent benchmarks on additional tasks and model pairs would further validate the claimed performance.