{"id":"arxiv:2510.03215","title":"Cache-to-Cache: Direct Semantic Communication Between Large Language Models","date":"2025-10-03","section":"LLMs: architectures, context, training","quality":0.499296,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/","source_url":"https://arxiv.org/abs/2510.03215","text_available":true,"authors":["Tianyu Fu","Zihan Min","Hanling Zhang","Jichao Yan"],"outgoing":[{"source":"arxiv:2510.03215","target":"arxiv:2407.21783","evidence":"title","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"}],"incoming":[{"source":"arxiv:2511.09149","target":"arxiv:2510.03215","evidence":"id","id":"arxiv:2511.09149","title":"Enabling Agents to Communicate Entirely in Latent Space","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/"},{"source":"arxiv:2605.22863","target":"arxiv:2510.03215","evidence":"title","id":"arxiv:2605.22863","title":"Latent Cache Flow: Model-to-Model Communication Without Text","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb6ed55f9fdbf8d212955b61ff74495737a2ecfacaa3645c7e4402b57e0c3e86/"},{"source":"arxiv:2608.03893","target":"arxiv:2510.03215","evidence":"title","id":"arxiv:2608.03893","title":"Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29465e80f45097327908d15926fd9b4a3c4c50431b2d525eb5aa9b92446fb2a/"},{"source":"arxiv:2608.11676","target":"arxiv:2510.03215","evidence":"title","id":"arxiv:2608.11676","title":"XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a1cbafd1249485a1818a70256d0d3ac5806d55b6900f005e9098c57d37918e44/"}],"text_parts":[{"start":0,"end":15977,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html"},{"start":15977,"end":31799,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html"},{"start":31799,"end":47725,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html"},{"start":47725,"end":58309,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html"}],"contents":[{"title":"Cache-to-Cache: Direct Semantic Communication Between Large Language Models","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"Abstract","offset":79,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"1 Introduction","offset":1543,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"2 Related Work","offset":5478,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"2.1 KV-Cache Sharing and Reuse","offset":5497,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"2.2 Multi-LLM Systems","offset":6598,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3 Method","offset":8842,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.1 Preliminaries","offset":8855,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.2 Oracles for Cache-to-Cache Communication","offset":10253,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.2.1 Cache Enrichment Oracle","offset":10692,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.2.2 Cache Transformation Oracle","offset":12754,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.3 C2C Design","offset":14142,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.3.1 Overview","offset":14162,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.3.2 Fuser Structure","offset":15427,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-1.html#text"},{"title":"3.3.3 Model Alignment","offset":16334,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"3.3.4 Training Scheme","offset":17073,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"4 Experiment","offset":17819,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"4.1 Setup","offset":17836,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"4.2 Main Results","offset":20744,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"4.3 Scaling Behavior","offset":22326,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"4.4 Ablation Study","offset":24534,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"4.5 Behavior Analysis","offset":25835,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"5 Discussion","offset":27613,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"6 Conclusion","offset":28962,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"Acknowledgments","offset":29509,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"Ethics Statement","offset":29878,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"Reproducibility Statement","offset":30016,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"Appendix A Appendix","offset":30492,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"A.1 Design Choice Exploration","offset":30516,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"A.1.1 Layer Alignment","offset":30620,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-2.html#text"},{"title":"A.1.2 Tokenization Alignment","offset":31984,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.1.3 Fuser Architecture","offset":34241,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.2 Additional Experimental Results","offset":35595,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.2.1 Cache Enrichment Detail","offset":35636,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.2.2 Strong-to-Weak Communication","offset":35915,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.2.3 Accuracy Breakdown","offset":36670,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.2.4 Progressive Behavior","offset":38616,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.3 Additional Experiment Setup","offset":39591,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.3.1 Cache Enrichment","offset":39628,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.3.2 Cache Transformation","offset":40622,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.3.3 Query-level routing","offset":41954,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.3.4 Evaluation Method","offset":42826,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.3.5 C2C Training","offset":44093,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.3.6 Evaluation Prompts","offset":46003,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.4 Additional Analysis","offset":46776,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.4.1 Effective Rank","offset":46805,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-3.html#text"},{"title":"A.4.2 Gating Behavior","offset":48044,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.4.3 Outlier cases in inference time","offset":49916,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.4.4 Example Model Output","offset":51149,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.4.5 Training Cost Analysis","offset":51947,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.4.6 Failure Mode Analysis","offset":52583,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.5 C2C Extensions and Future Work","offset":53569,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.5.1 One Receiver, Multiple Sharers","offset":53949,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.5.2 Multiple Receivers Multiple Sharers","offset":54446,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"A.5.3 Incorporating with Agentic Flow","offset":56462,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"},{"title":"Instructions for reporting errors","offset":57345,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/part-4.html#text"}],"text_code_points":58309,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b5a2a7fccb7e223ec3c3769c18189f6a6e4c18e2b074567db94cd7db2450723b/source.md"}