{"id":"arxiv:2605.22863","title":"Latent Cache Flow: Model-to-Model Communication Without Text","date":"2026-05-19","section":"LLMs: architectures, context, training","quality":-0.148425,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/","source_url":"https://arxiv.org/abs/2605.22863","text_available":true,"authors":["Maximillian Rossi","Prajwal Raghunath","Eugene Wu"],"outgoing":[{"source":"arxiv:2605.22863","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":15844,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html"},{"start":15844,"end":31745,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html"},{"start":31745,"end":37344,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html"}],"contents":[{"title":"Latent Cache Flow: Model-to-Model Communication Without Text","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"Abstract","offset":64,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"Keywords:","offset":1240,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"1 Introduction","offset":1258,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"2 Background: Cache-to-Cache","offset":3598,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"3 Latent Cache Flow","offset":5549,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"3.1 Latent Cache Flow for Exact Context Sharing","offset":5858,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"3.2 LCF-X: Cross-Context Communication","offset":9290,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"4 Experiments","offset":11706,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"4.1 Shared-Context","offset":11954,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"4.1.1 Experimental Setup","offset":11978,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"4.1.2 Training Cost","offset":13580,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"4.1.3 Shared-Context Results","offset":14952,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-1.html#text"},{"title":"4.1.4 Layer Pruning","offset":16412,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"4.2 Cross-Context Information Transfer","offset":17393,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"4.2.1 Experimental Setup","offset":17437,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"4.2.2 Cross-Context Results","offset":19990,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"5 Limitations and Future Work","offset":21606,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"6 Conclusion","offset":22326,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Appendix A Full Results Overview","offset":23528,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Appendix B Training Details and C2C Comparison Protocol","offset":24530,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"B.1 Matched Training Protocol","offset":25316,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Training cost.","offset":25351,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Why 300 steps.","offset":25839,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"B.2 Eval-Loss Convergence","offset":26394,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"B.3 Training-Curve Summary","offset":26989,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Appendix C Statistical Significance","offset":27405,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"C.1 Main LCF Variants","offset":28358,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"C.2 Pruned LCF-128 Variants","offset":28749,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Appendix D Layer Pruning","offset":29149,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Parameter accounting.","offset":29873,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Pruning method.","offset":30252,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Why pruning works.","offset":31719,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-2.html#text"},{"title":"Limitations.","offset":32096,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html#text"},{"title":"Appendix E LCF-X Training Details","offset":32338,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html#text"},{"title":"E.1 Hyperparameters","offset":33112,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html#text"},{"title":"E.2 Training Trajectory","offset":33743,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html#text"},{"title":"E.3 Architectural Cost","offset":34323,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html#text"},{"title":"E.4 Span-Scheme Invariance","offset":35344,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html#text"},{"title":"Instructions for reporting errors","offset":36380,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/part-3.html#text"}],"text_code_points":37344,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/source.md"}