Add Cache-to-Cache - Direct Semantic Communication Between LLMs

[minzh23]

Summary

This PR adds Cache-to-Cache, a novel research work on direct semantic communication between Large Language Models via KV-Cache.

What is Cache-to-Cache?

Cache-to-Cache introduces a new paradigm for inter-LLM communication that goes beyond traditional text-based approaches. Instead of forcing models to communicate through token sequences, Cache-to-Cache enables direct semantic transfer via KV-Cache manipulation.

Key Features

• Direct Semantic Communication: LLMs communicate through KV-Cache instead of text, preserving rich semantic information
• Eliminates Generation Latency: Avoids token-by-token generation overhead
• Neural Cache Fusion: Uses a neural network to project and fuse KV-caches between models
• Superior Performance: Achieves significant improvements over both individual models and text-based communication

Category

Multi-agent, Communication

Changes Made

• Added Cache-to-Cache entry in alphabetical order (between "bumpgen" and "Cal.ai")
• Included comprehensive description of the research contribution
• Added complete links: Project Page, Paper (arXiv), GitHub repository, and HuggingFace collection

Resources

• 🌐 Project Page
• 📑 Paper
• 💻 GitHub
• 🤗 HuggingFace