J-Comm: Jacobian-Guided Latent Communication between LLM Agents

A new arXiv paper (2610.32871v1) presents J-Comm, a framework for latent communication between LLM agents that combines a training-free Geometric Agreement for Latent Reasoning (GEAR) method with an H+J (Hidden-and-Jacobian) interface. In Qwen3-8B to Qwen3-4B communication, J-Comm posts the highest mean score across seven benchmarks, including 94.80% on MATH500, 54.55% on GPQA-Diamond, 73.33% on AIME2024+2025, and 62.93% success on ALFWorld, beating TextMAS by 1.29-8.59 percentage points while cutting generated text tokens by 19.2-74.8%. The approach lets a Receiver reconstruct the Sender's cues using its own dictionary and inject them into intermediate layers during prefill, avoiding shared key-value caches or identical model widths.

arXiv:2610.32871v1 Announce Type: new Abstract: Latent communication requires both an effective continuous reasoning process and an interface through which an external model can leverage its intermediate computation. We present J-Comm, a framework comprising two core components. First, a training-free Geometric Agreement for Latent Reasoning GEAR reweights token candidates over existing model embeddings, which monotonically optimizes geometric consensus among candidate tokens in embedding space at each step. Second, we introduce an H+J Hidden-and-Jacobian communication interface that gives the Receiver two complementary views of the Sender's reasoning. The H channel supplies continuous inputs derived from feedback distributions, while the J channel conveys intermediate-state cues in vocabulary-indexed Jacobian-lens coordinates. Intuitively, H provides inputs for the Receiver to reason from, while J offers additional guidance within its computation. The Receiver reconstructs these cues using its own dictionary and injects them into intermediate layers during prefill, enabling communication without sharing key-value caches or requiring identical model widths. In Qwen3-8B to Qwen3-4B communication, J-Comm achieves the highest mean score across seven benchmarks, including 94.80% on MATH500, 54.55% on GPQA-Diamond, 73.33% on AIME2024+2025, and 62.93% success on ALFWorld. Compared with TextMAS, it improves performance by 1.29-8.59 percentage points and reduces generated text tokens by 19.2-74.8%.