J-Comm preprint reports agent performance gains with fewer generated text tokens

Researchers report in a preprint that J-Comm, a framework for communication between language models, improves benchmark performance by 1.29-8.59 percentage points while reducing generated text tokens by 19.2-74.8% compared with TextMAS. The framework lets a receiving model use intermediate reasoning signals from a sending model. It reconstructs those signals using the receiver's own dictionary and inserts them into intermediate processing layers, without requiring identical model widths or shared key-value caches. In tests connecting Qwen3-8B to Qwen3-4B, J-Comm achieved the highest mean score across seven benchmarks. The reported token reduction concerns generated text; the interface also transfers intermediate reasoning signals.