Cache-to-Cache: Direct Semantic Communication Between LLMs (2025)
The paper proposes passing model state directly between LLMs instead of translating it into intermediate text.
Cache-to-Cache uses a neural network to project and fuse one model’s KV-cache into another model’s cache, with a learned gate choosing which target layers use the transfer. The authors argue this preserves richer internal semantics and avoids token-by-token communication overhead. In their experiments, C2C beats individual models by 6.4-14.2% average accuracy, exceeds text-based communication by about 3.1-5.4%, and reports a 2.5x average latency speedup. The paper is marked published in ICLR ’26. HN · ArXiv's note
Cache-to-Cache uses a neural network to project and fuse one model’s KV-cache into another model’s cache, with a learned gate choosing which target layers use the transfer. The authors argue this preserves richer internal semantics and avoids token-by-token communication overhead. In their experiments, C2C beats individual models by 6.4-14.2% average accuracy, exceeds text-based communication by about 3.1-5.4%, and reports a 2.5x average latency speedup. The paper is marked published in ICLR ’26. HN · ArXiv's note
score 5