Tsinghua's C2C Lets LLMs Skip Text and Merge KV-Caches Directly, 2.5x Faster with +14.2% Accuracy

anselm · x · 2026-09-19

Tsinghua and Infinigence teams open-sourced C2C (Cache-to-Cache), an ICLR 2026 paper that removes text from multi-agent LLM communication: a lightweight Neural Fuser rotates and grafts model A's KV-Cache directly into model B's, with a learnable gate deciding which layers absorb external cache. Results: 2.0–2.5x faster inference, up to 14.2% accuracy gain over a single model, and 5% over text-based agent collaboration — challenging the assumption that LLMs must talk in human language.

Related event: Tsinghua Open-Sources C2C: LLMs Communicate via KV-Cache, 2.5x Faster Inference(2 posts)→

Original post →

More from coding & agent

coding & agent channel →