Cache-to-Cache lets multi-LLM systems share KV caches instead of text handoffs

Researchers from Tsinghua University and other Chinese institutions proposed Cache-to-Cache (C2C), letting multi-LLM systems exchange data via key-value caches instead of text. Experiments showed 3.1 to 5.4 percentage point accuracy gains over text handoffs and lower latency across tested model pairs. Published at ICLR 2026, the team plans a September release of an agent-managed KV-Cache implementation and serving system.