Bypassing Natural Language for LLM Agents

Summary & Methodology Analysis

Large language model based multi-agent systems normally communicate using natural language, which introduces a compression bottleneck that drops continuous information contained in hidden representations (the internal continuous vector states inside a transformer, which is the underlying neural network architecture powering modern language models). This paper investigates whether off-the-shelf LLM agents can bypass natural language entirely and communicate through continuous representations using training-free alignment alone. The method retains only the final K hidden states of the generated message tokens from the sender agent, obtains reference embeddings by looking up the resulting message tokens in the token embedding matrix, centers both the sender hidden states and reference embeddings to remove global offsets, and whitens both sets of points to rescale principal directions. It then solves an orthogonal Procrustes problem (a mathematical technique to find an optimal rotation matrix between two sets of points) in closed-form using singular value decomposition to find the optimal rotation matrix that aligns whitened sender states to reference embeddings, restores the overall scale pattern and location of the reference embeddings, calibrates each aligned vector to match the typical norm of vocabulary embeddings, and moves each calibrated vector slightly toward its nearest vocabulary embedding under cosine similarity via vocabulary anchoring. Finally, it prepends the resulting continuous aligned prefix directly to the receiver agent prompt at the embedding layer. To evaluate this approach, the paper uses Qwen3, OLMo3-7B-Think, GSM8K, AIME24, AIME25, GPQA-Diamond, ARC-Challenge, MedQA, MBPP+, and HumanEval+. StateBridge achieves the best or tied-best score on 22 out of 26 model-task pairs, consistently outperforming the strongest baseline, and achieves the best average score in all four model settings, improving over the best baseline by 2.4 to 2.9 points. On OLMo3-7B-Think, LatentMAS averages 55.1% while StateBridge reaches 76.7%. Ablations show that replacing Procrustes with ridge regression results in average performance drops from 82.4% to 74.9% on Qwen3-4B, and replacing the aligned prefix with random noise drops average performance to 48.8%. Regarding limitations, on GSM8K for Qwen3 models, StateBridge trails the best baseline, likely because continuous prefixes affect output formatting under exact-match evaluation. Other computational considerations, memory limits, and hardware requirements are not specified by the paper.