When LLMs communicate without intermediate text
Cache-to-Cache introduces a new way for LLMs to communicate: models transfer internal representations directly through their KV caches, instead of first writing messages to each other.
Tianyu Fu and co-authors introduce Cache-to-Cache (C2C), a direct exchange of internal context representations between language models. A trained module translates and combines their KV caches – stored states used in attention computation. This removes the need to generate intermediate messages token by token. [1]
The study reports higher accuracy and an average 2.5-fold speedup over text communication. The approach requires access to model internals; incorrect information can propagate. Scaling to many models remains unresolved. [2]
Our interpretation: this example shifts attention from individual chatbots to cooperating models. Their collaboration could increasingly operate through technical representations.
Sources
This indication by Martin Ebner is licensed under
CC BY 4.0
Comments 0
Comments are moderated. New comments appear after review by the editors.
No comments yet.