Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse

arxiv:2608.03893

Original source · Paper JSON

LLMs: architectures, context, training · 2026-08-04

Cites (3) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

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

Contents

• Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse
• Abstract
• 1 Introduction
• Contributions.
• 2 Background and motivation
• 2.1 Problem formulation
• 2.2 Prior approaches
• 2.3 Linear structure in cross-model KV
• How many source layers does the mapper need?
• 3 Mapper design
• 3.1 Per-head ridge regression
• Fitting.
• 3.2 Cross-layer source selection
• 3.3 Content-space mapping (RoPE factoring)
• 4 Experiments
• 4.1 Setup
• Model families.
• Benchmarks.
• Mapper configuration.
• 4.2 Main results
• 4.3 Ablation
• 4.4 Substituting MLP for ridge
• 4.5 What determines transfer quality?
• Attention-output cosine.
• Where the mapper’s error lands.
• MLP intervention.
• 4.6 Multi-turn handoff
• 4.7 Prefill latency
• 5 Broader impact, future work, and limitations
• Broader impact.
• Future work.
• Limitations.
• 6 Conclusion
• Appendix A Cross-model KV cache transfer methods
• Appendix B Linear structure analysis
• Source layer selection by greedy forward selection.
• Appendix C Extended ablation
• Source layer count.
• Sequential component removal.
• Calibration protocol sensitivity.
• Calibration domain across benchmarks.
• Appendix D Experimental setup details
• Benchmarks.
• Coverage.
• Perplexity protocol.
• Mapper configuration.
• Serving cost.
• Appendix E Run information
• Appendix F Extended main results
• Floor-normalized retention.
• Appendix G Prefill latency details
• Setup.
• Caveats.
• Appendix H Source-layer count selection
• Protocol.
• Leave-one-benchmark-out re-selection.
• Held-out benchmarks.
• Instructions for reporting errors