StateBridge Aligns Agent Hidden States with a Closed-Form Rotation, Bypassing the Text Communication Bottleneck Without Training

Multi-agent systems typically encode planning, critique, and revision results as text before passing them to the next agent. When this information is sampled into tokens, confidence signals and alternative paths contained in the sender’s hidden states are compressed away. Instead, StateBridge extracts the final-layer hidden states of the message’s last 64 tokens and treats them as a point cloud. Through centering, whitening, and orthogonal Procrustes alignment, it derives a closed-form rotation that preserves angles and distances. It then calibrates the vector norms and nudges each vector slightly toward its nearest vocabulary embedding before prepending the result to the receiving agent’s input as a continuous prefix—all without updating the model weights. The [paper](https://arxiv.org/abs/2608.13317) has been accepted to [COLM 2026](https://colmweb.org/AcceptedPapers.html), and the implementation is publicly available on [GitHub](https://github.com/YanwenPneg/StateBridge).

The authors evaluated a four-agent Planner–Critic–Refiner–Judger pipeline using Qwen3 4B, 8B, and 32B, as well as OLMo3-7B-Think, across mathematics, question answering, and code generation. StateBridge achieved or tied for the best result in 22 of 26 model–task combinations, improving each model’s average score by 2.4 to 2.9 points over the strongest baseline. Notable examples include Qwen3-8B, where GPQA rose from the strongest baseline’s 45.5% to 52.5% and AIME24 increased from 56.7% to 63.3%. On Qwen3-32B, GPQA improved from 58.3% to 64.1%. Unlike transferring KV caches across layers, StateBridge operates only through the input embedding interface, so it did not suffer the same degree of performance collapse when applied to the different OLMo architecture.

The limitations are equally concrete. Agents in the experiments shared the same model weights, and the work has not yet shown that latent representations can be transferred directly between different model families. The method also requires a forward hook to access hidden states, making it largely unusable with closed APIs. Its alignment procedure includes an approximately O(d³) decomposition and an O(KVd) vocabulary search, while production throughput and cross-machine transfer have not been tested. Some GSM8K configurations also underperformed the text baseline because of formatting and other issues. Key engineering questions include whether continuous messages can preserve their semantics across quantization, tensor parallelism, and heterogeneous models—and whether unreadable agent communication will make debugging, auditing, and safety monitoring more difficult.