A new arXiv paper proposes letting large language models swap their internal cache states instead of talking to each other only in plain text.
The paper, titled "Cache-to-Cache: Direct Semantic Communication Between LLMs," was posted to arXiv and has been circulating in developer discussions, gathering modest engagement so far. The core idea: when one model currently needs another model's help, it generates text, and the second model has to read that text and re-encode it into its own internal representation before it can use it. The authors propose skipping that round trip by passing the KV-cache - the model's internal semantic state - directly between models instead.
This matters because multi-model pipelines are becoming a default architecture, not a novelty. Chains of LLMs calling other LLMs, or agent frameworks that hand tasks between specialized models, all currently rely on text as the connective tissue. Text is legible but wasteful: generating it costs compute, and reading it back in loses some of the nuance the first model actually had. A cache-level interface is a bet that speed and fidelity matter more than having a human-readable log of what one model told another.
That tradeoff cuts both ways. The same plain-text interface this approach treats as overhead is also the only place a developer can currently see what's going wrong when a multi-model system fails. Skip it, and debugging gets harder even as throughput improves. With only 55 points and 11 comments in early discussion, this is still a niche research idea, not a shipped feature - worth watching, not worth building around yet.