Latent Cache Flow: Model-to-Model Communication Without Text
A paper proposes Latent Cache Flow, a way for LLMs to share information through compressed cache transfers instead of text. The authors say it is smaller, faster, and more accurate than prior cache exchange methods in early tests.
Intelligence analysis by GPT-5.4 Mini
The paper argues that text is a slow and lossy way for AI agents to communicate, because one model must decode and another must re-encode the same information. Latent Cache Flow aims to move summarized cache state directly between models, shrinking the adapter and improving performance when contexts differ.
Two computer brains usually pass messages by writing words to each other. That works, but it is slow and can drop some of the meaning, like trying to copy a picture by describing it aloud.
This paper suggests a different trick. Instead of turning everything into text, one brain sends a compact hidden package of what it knows, so the other brain can use it more directly.
The authors say this can be smaller and faster than older methods. It is a bit like handing someone a folded map instead of asking them to listen to a long street-by-street explanation.
Analysis
What the paper is proposing
The paper starts from a simple complaint: LLM agents usually talk through text, and that adds latency plus information loss. The authors point to a newer line of work, Cache-to-Cache, which tries to exchange KV caches by learning adapters that translate one model’s cache into another’s. Their critique is that those adapters are large, expensive to train, and rely on the two models having the same context, which makes them awkward for real agent communication.
Latent Cache Flow
Latent Cache Flow (LCF) is the authors’ answer to that problem. They say the method jointly translates and compresses keys and values, which reduces the adapter to about 4% of the size of C2C. They also change the communication goal: instead of trying to mirror individual tokens, the adapter is designed to send a summary of new information that the receiver model does not already have. That is meant to fit situations where two models have different contexts.
Reported early results
The abstract gives two headline comparisons from early experiments. In shared-context settings, the authors say a 13 MB LCF adapter can be more accurate than a 956 MB C2C adapter. In different-context settings, they report LCF as 23% more accurate and 8.5x faster than text-based communication. The submission is short, at 6 pages and 5 figures, so this reads as an early-stage research result rather than a finished system claim.
Key points
- The paper argues that text-based AI communication is slow and can lose information.
- Latent Cache Flow sends compressed cache information instead of converting everything into words.
- The authors say the adapter is about 4% the size of a prior cache-exchange method.
- They report early gains in both shared-context and different-context settings.
- The work is presented as an early research paper, not a deployed product.
