Cache-to-Cache: Direct Semantic Communication Between Large Language Models

Researchers from Meta AI and the University of California, Berkeley have introduced a novel approach called Cache-to-Cache: Direct Semantic Communication Between Large Language Models. This method enables direct communication between two large language models without the need for intermediate tokenization or encoding. The researchers used a transformer-based neural network architecture to develop a cache-based system that stores and retrieves information from both models. The system demonstrates improved efficiency and speed in transferring information between models, with a significant reduction in the number of tokens required for communication. The authors propose that this technique can be applied to various applications, including natural language processing, dialogue systems, and knowledge graphs. The paper is currently available on arXiv and has received 12 points on Hacker News.