J-Comm links LLM agents with hidden and Jacobian latent channels
Original title: J-Comm: Jacobian-Guided Latent Communication between LLM Agents
Qwen3-8B to Qwen3-4B communication scored highest across seven benchmarks and cut generated text tokens by 19.2–74.8% versus TextMAS.
J-Comm combines training-free Geometric Agreement for Latent Reasoning (GEAR), which reweights token candidates over existing embeddings, with an H+J interface. The H channel sends continuous inputs from feedback distributions; the J channel sends vocabulary-indexed Jacobian-lens cues. The receiver rebuilds those cues with its own dictionary and injects them during prefill, without shared key-value caches or matching model widths. In Qwen3-8B to Qwen3-4B tests, J-Comm led seven benchmarks, including 94.80% on MATH500, 54.55% on GPQA-Diamond, 73.33% on AIME2024+2025 and 62.93% on ALFWorld. Versus TextMAS it gained 1.29–8.59 points and used 19.2–74.8% fewer generated text tokens.
- GEAR is training-free and reweights candidates over existing embeddings.
- Communication does not share key-value caches or require identical widths.
- MATH500 score was 94.80% and GPQA-Diamond was 54.55%.
- AIME2024+2025 was 73.33% and ALFWorld success was 62.93%.
- Gains over TextMAS were 1.29 to 8.59 percentage points.
From the original article
1 min read at arxiv.orgLatent 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.
By Xiangwei Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.