Cache-to-Cache: LLM Agents Share Context Without Text

A team of researchers published a paper about a method called Cache-to-Cache, where they are able to share the internal KV cache amongst language models instead of sharing text data. They were able to get up to 6.4–14.2% gains on various benchmarks and up to ~2.5x speedups! Accepted to ICLR 2026 :)

Currently, the only way for agents to communicate context is via serialization to text, which is pretty token heavy. Using Cache-to-Cache allows you to share the internal state without having to serialize to text. It’s still just a research project right now, so it doesn’t exist anywhere yet, but once it gets turned into a product, I think it could potentially speed up and lower costs in your agent pipelines.