Latent Cache Flow: Model-to-Model Communication Without Text

arxiv:2605.22863

Original source · Paper JSON

LLMs: architectures, context, training · 2026-05-19

Cites (1) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

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

Contents

• Latent Cache Flow: Model-to-Model Communication Without Text
• Abstract
• Keywords:
• 1 Introduction
• 2 Background: Cache-to-Cache
• 3 Latent Cache Flow
• 3.1 Latent Cache Flow for Exact Context Sharing
• 3.2 LCF-X: Cross-Context Communication
• 4 Experiments
• 4.1 Shared-Context
• 4.1.1 Experimental Setup
• 4.1.2 Training Cost
• 4.1.3 Shared-Context Results
• 4.1.4 Layer Pruning
• 4.2 Cross-Context Information Transfer
• 4.2.1 Experimental Setup
• 4.2.2 Cross-Context Results
• 5 Limitations and Future Work
• 6 Conclusion
• Appendix A Full Results Overview
• Appendix B Training Details and C2C Comparison Protocol
• B.1 Matched Training Protocol
• Training cost.
• Why 300 steps.
• B.2 Eval-Loss Convergence
• B.3 Training-Curve Summary
• Appendix C Statistical Significance
• C.1 Main LCF Variants
• C.2 Pruned LCF-128 Variants
• Appendix D Layer Pruning
• Parameter accounting.
• Pruning method.
• Why pruning works.
• Limitations.
• Appendix E LCF-X Training Details
• E.1 Hyperparameters
• E.2 Training Trajectory
• E.3 Architectural Cost
• E.4 Span-Scheme Invariance
• Instructions for reporting errors