代理系統
StateBridge Transmits Agents’ Hidden States via Closed-Form Alignment, Topping or Tying 22 of 26 Test Configurations
StateBridge uses an orthogonal Procrustes transformation to map the sending agent’s final-layer hidden states into the receiving agent’s input embedding space, without training a projector. Its average scores across four models exceeded the strongest baseline by 2.4 to 2.9 points, but it currently supports only agents using identical model weights.
Multi-agent LLM systems typically write intermediate reasoning as text and pass it to the next agent. This compresses continuous hidden states into discrete tokens, leaving the receiver able to recover only the token embeddings. StateBridge instead captures the sender’s final-layer hidden states as it generates a message and uses the corresponding token embeddings as references. After centering, whitening, orthogonal Procrustes alignment, norm calibration, and vocabulary anchoring, it injects the states into the next agent as a continuous prefix through `inputs_embeds`. The orthogonal mapping preserves distances and angles between states while avoiding the need to retrain a projector for each model or task.
The authors evaluated Qwen3-4B, Qwen3-8B, Qwen3-32B, and OLMo3-7B-Think in a four-agent Planner–Critic–Refiner–Judger pipeline across eight benchmarks, including GPQA, MedQA, AIME, GSM8K, and HumanEval+. StateBridge achieved the best or tied-best result in 22 of 26 model–task combinations. Across all four model configurations, its average score surpassed the strongest text-based or KV-cache baseline by 2.4 to 2.9 points; on GPQA, Qwen3-8B improved by 7.0 points. On OLMo3, LatentMAS, which injects the KV cache layer by layer, averaged only 55.1%, compared with 76.7% for StateBridge. This suggests that relying solely on the shared input-embedding interface makes the method less dependent on a Transformer’s internal architecture.
This is not a general-purpose cross-model agent protocol: the experiments assume that every agent shares the same set of weights, the prefix length is fixed, and performance declined on some GSM8K configurations. The tests also used stochastic decoding with a temperature of 0.6, and the current findings come primarily from a single implementation by the authors, without independent replication. From an engineering perspective, future work should examine its effects on heterogeneous models, batched serving, KV-cache management, and end-to-end latency—not just task accuracy.