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

arxiv:2510.03215

Original source · Paper JSON

LLMs: architectures, context, training · 2025-10-03

Cites (1) · Cited by (4)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• Cache-to-Cache: Direct Semantic Communication Between Large Language Models
• Abstract
• 1 Introduction
• 2 Related Work
• 2.1 KV-Cache Sharing and Reuse
• 2.2 Multi-LLM Systems
• 3 Method
• 3.1 Preliminaries
• 3.2 Oracles for Cache-to-Cache Communication
• 3.2.1 Cache Enrichment Oracle
• 3.2.2 Cache Transformation Oracle
• 3.3 C2C Design
• 3.3.1 Overview
• 3.3.2 Fuser Structure
• 3.3.3 Model Alignment
• 3.3.4 Training Scheme
• 4 Experiment
• 4.1 Setup
• 4.2 Main Results
• 4.3 Scaling Behavior
• 4.4 Ablation Study
• 4.5 Behavior Analysis
• 5 Discussion
• 6 Conclusion
• Acknowledgments
• Ethics Statement
• Reproducibility Statement
• Appendix A Appendix
• A.1 Design Choice Exploration
• A.1.1 Layer Alignment
• A.1.2 Tokenization Alignment
• A.1.3 Fuser Architecture
• A.2 Additional Experimental Results
• A.2.1 Cache Enrichment Detail
• A.2.2 Strong-to-Weak Communication
• A.2.3 Accuracy Breakdown
• A.2.4 Progressive Behavior
• A.3 Additional Experiment Setup
• A.3.1 Cache Enrichment
• A.3.2 Cache Transformation
• A.3.3 Query-level routing
• A.3.4 Evaluation Method
• A.3.5 C2C Training
• A.3.6 Evaluation Prompts
• A.4 Additional Analysis
• A.4.1 Effective Rank
• A.4.2 Gating Behavior
• A.4.3 Outlier cases in inference time
• A.4.4 Example Model Output
• A.4.5 Training Cost Analysis
• A.4.6 Failure Mode Analysis
• A.5 C2C Extensions and Future Work
• A.5.1 One Receiver, Multiple Sharers
• A.5.2 Multiple Receivers Multiple Sharers
• A.5.3 Incorporating with Agentic Flow
• Instructions for reporting errors