Cache-to-Cache: Direct Semantic Communication Between LLMs (2025)

Research paper proposes Cache-to-Cache, letting LLMs communicate directly via KV-cache semantics instead of tokenized text.

An arXiv paper (2510.03215) introduces Cache-to-Cache (C2C), a method for direct semantic communication between large language models using their KV caches rather than text tokens. A learnable projection transfers one model's cached semantics into another model's latent space, reportedly improving response quality while cutting the tokenization and decoding overhead of text-based pipelines. The work targets more efficient multi-LLM and agentic systems and drew 42 points and 8 comments on Hacker News.

• Proposes LLM-to-LLM communication in latent space via KV caches.
• Avoids tokenization and decoding overhead of text-based agent pipelines.
• Uses a learnable projection to transfer semantics between different models.
• Trended on Hacker News with 42 points and 8 comments.