AI TrackerFrontier desk · AEST
Back to news
normalaitech

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

arXiv cs.CL · 24 Oct 2026, 02:42 AEST

J-Comm is a framework designed for latent communication between LLM agents, featuring a training-free Geometric Agreement for Latent Reasoning and a Hidden-and-Jacobian communication interface. It has demonstrated superior performance on multiple benchmarks compared to previous methods.

Key facts

Read aloud

J-Comm is a framework designed for latent communication between L L M agents, featuring a training-free Geometric Agreement for Latent Reasoning and a Hidden-and-Jacobian communication interface. It has demonstrated superior performance on multiple benchmarks compared to previous methods. 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. In Qwen3-8B to Qwen3-4B communication, J-Comm achieves the highest mean score across seven benchmarks, including ninety four point eight zero percent on MATH500, fifty four point five five percent on GPQA-Diamond, seventy three point three three percent on AIME2024+twenty twenty five, and sixty two point nine three percent success on ALFWorld. Compared with TextMAS, it improves performance by one point two nine-eight point five nine percentage points and reduces generated text tokens by nineteen point two-seventy four point eight percent.

Sources in this cluster