{"id":"arxiv:2608.03893","title":"Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse","date":"2026-08-04","section":"LLMs: architectures, context, training","quality":0.145329,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/","source_url":"https://arxiv.org/abs/2608.03893","text_available":true,"authors":["Taekyung Heo","Rasoul Shafipour","Ritchie Zhao","Maximilian Golub"],"outgoing":[{"source":"arxiv:2608.03893","target":"arxiv:2405.07987","evidence":"title","id":"arxiv:2405.07987","title":"The Platonic Representation Hypothesis","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6c7695929f8bbcea43ee98da9cd8f2fe806f2ba292f0b4e3e7fea2b64254afb4/"},{"source":"arxiv:2608.03893","target":"arxiv:2407.21783","evidence":"id","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"},{"source":"arxiv:2608.03893","target":"arxiv:2510.03215","evidence":"title","id":"arxiv:2510.03215","title":"Cache-to-Cache: Direct Semantic Communication Between Large Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/"}],"incoming":[],"text_parts":[{"start":0,"end":14469,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html"},{"start":14469,"end":30351,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html"},{"start":30351,"end":46117,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html"},{"start":46117,"end":48975,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-4.html"}],"contents":[{"title":"Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"Abstract","offset":97,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"1 Introduction","offset":1602,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"Contributions.","offset":5111,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"2 Background and motivation","offset":6358,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"2.1 Problem formulation","offset":6390,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"2.2 Prior approaches","offset":8317,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"2.3 Linear structure in cross-model KV","offset":9444,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"How many source layers does the mapper need?","offset":12338,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"3 Mapper design","offset":12972,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"3.1 Per-head ridge regression","offset":13564,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"Fitting.","offset":14454,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-1.html#text"},{"title":"3.2 Cross-layer source selection","offset":16266,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"3.3 Content-space mapping (RoPE factoring)","offset":17629,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4 Experiments","offset":19215,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4.1 Setup","offset":19603,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"Model families.","offset":19618,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"Benchmarks.","offset":19998,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"Mapper configuration.","offset":21012,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4.2 Main results","offset":21790,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4.3 Ablation","offset":22978,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4.4 Substituting MLP for ridge","offset":23962,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4.5 What determines transfer quality?","offset":25414,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"Attention-output cosine.","offset":26367,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"Where the mapper’s error lands.","offset":27526,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"MLP intervention.","offset":28190,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4.6 Multi-turn handoff","offset":29361,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"4.7 Prefill latency","offset":30326,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-2.html#text"},{"title":"5 Broader impact, future work, and limitations","offset":31042,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Broader impact.","offset":31093,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Future work.","offset":31344,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Limitations.","offset":32769,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"6 Conclusion","offset":33775,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix A Cross-model KV cache transfer methods","offset":34865,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix B Linear structure analysis","offset":35534,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Source layer selection by greedy forward selection.","offset":35967,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix C Extended ablation","offset":37566,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Source layer count.","offset":37819,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Sequential component removal.","offset":38432,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Calibration protocol sensitivity.","offset":39128,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Calibration domain across benchmarks.","offset":39684,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix D Experimental setup details","offset":40372,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Benchmarks.","offset":40724,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Coverage.","offset":41048,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Perplexity protocol.","offset":41488,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Mapper configuration.","offset":41960,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Serving cost.","offset":42423,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix E Run information","offset":43559,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix F Extended main results","offset":43750,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Floor-normalized retention.","offset":44053,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix G Prefill latency details","offset":44746,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Setup.","offset":44915,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Caveats.","offset":45457,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Appendix H Source-layer count selection","offset":45842,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Protocol.","offset":46101,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-3.html#text"},{"title":"Leave-one-benchmark-out re-selection.","offset":46921,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-4.html#text"},{"title":"Held-out benchmarks.","offset":47485,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-4.html#text"},{"title":"Instructions for reporting errors","offset":48011,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/part-4.html#text"}],"text_code_points":48975,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/source.md"}