J-Comm
A latent communication framework for LLM-based multi-agent systems that transmits intermediate reasoning states through Jacobian-guided coordinates, enabling heterogeneous models to exchange computational information efficiently without sharing KV caches.
Timeline (last 90 days)
J-Comm proposes Jacobian-guided latent communication that lets a receiver model directly exploit the sender's intermediate computation without sharing KV caches or requiring equal width.
J-Comm consists of two components: GEAR (Geometric Agreement for Latent Reasoning) and the H+J communication interface.
In the Qwen3-8B to Qwen3-4B communication setting, J-Comm achieves the highest average score across seven benchmarks.
J-Comm reaches 94.80% on MATH500, 54.55% on GPQA-Diamond, 73.33% on AIME2024+2025, and a 62.93% success rate on ALFWorld.
Compared with TextMAS, J-Comm's performance gains range from 1.29 to 8.59 percentage points while generating 19.2% to 74.8% fewer text tokens.
J-Comm's current experiments focus on cross-scale communication within the Qwen3 family (8B to 4B); communication effectiveness at larger scales and across model families remains to be verified.
J-Comm does not modify model weights and only injects signals at inference time, which differs from directions such as model merging and representation alignment.
All Facts (7)
J-Comm proposes Jacobian-guided latent communication that lets a receiver model directly exploit the sender's intermediate computation without sharing KV caches or requiring equal width.
50%UnverifiedJ-Comm consists of two components: GEAR (Geometric Agreement for Latent Reasoning) and the H+J communication interface.
50%UnverifiedIn the Qwen3-8B to Qwen3-4B communication setting, J-Comm achieves the highest average score across seven benchmarks.
50%UnverifiedJ-Comm reaches 94.80% on MATH500, 54.55% on GPQA-Diamond, 73.33% on AIME2024+2025, and a 62.93% success rate on ALFWorld.
50%UnverifiedCompared with TextMAS, J-Comm's performance gains range from 1.29 to 8.59 percentage points while generating 19.2% to 74.8% fewer text tokens.
50%UnverifiedJ-Comm's current experiments focus on cross-scale communication within the Qwen3 family (8B to 4B); communication effectiveness at larger scales and across model families remains to be verified.
50%UnverifiedJ-Comm does not modify model weights and only injects signals at inference time, which differs from directions such as model merging and representation alignment.
50%