{"id":"arxiv:2511.09149","title":"Enabling Agents to Communicate Entirely in Latent Space","date":"2025-11-12","section":"LLMs: architectures, context, training","quality":0.297245,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/","source_url":"https://arxiv.org/abs/2511.09149","text_available":true,"authors":["Zhuoyun Du","Runze Wang","Huiyu Bai","Zouying Cao"],"outgoing":[{"source":"arxiv:2511.09149","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/"},{"source":"arxiv:2511.09149","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2412.06769","title":"Training Large Language Models to Reason in a Continuous Latent Space","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/"},{"source":"arxiv:2511.09149","target":"arxiv:2501.14082","evidence":"id","id":"arxiv:2501.14082","title":"Communicating Activations Between Language Model Agents","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/"},{"source":"arxiv:2511.09149","target":"arxiv:2502.05171","evidence":"id","id":"arxiv:2502.05171","title":"Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/24a6ab865777290627cecebd3f06ea3daef01622f102d32a95f7fb2a5a42880f/"},{"source":"arxiv:2511.09149","target":"arxiv:2510.03215","evidence":"id","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/"},{"source":"arxiv:2511.09149","target":"url:anthropic.com/research/reasoning-models-dont-say-think","evidence":"id","id":"url:anthropic.com/research/reasoning-models-dont-say-think","title":"Reasoning Models Don't Always Say What They Think","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4ca0654119289c784e751677f3ce63742c77f58afe54849ad74eafb2d1bccb8a/"}],"incoming":[],"text_parts":[{"start":0,"end":15755,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html"},{"start":15755,"end":31379,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html"},{"start":31379,"end":47254,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html"},{"start":47254,"end":57720,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html"}],"contents":[{"title":"Enabling Agents to Communicate Entirely in Latent Space","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Abstract","offset":59,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"1 Introduction","offset":1629,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"2 Related Work","offset":5430,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Latent Reasoning in LLMs.","offset":5449,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Multi-agent Communication.","offset":6485,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"3 Interlat","offset":7865,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"3.1 Latent Communication","offset":8857,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"3.2 Training Procedure","offset":10168,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Conditional thought separation.","offset":11350,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Plan-aligned regulation.","offset":12104,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Curriculum Learning.","offset":13458,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"4 Information Compression","offset":14165,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Compression via Latent-Space Reasoning.","offset":14771,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"Training Objective.","offset":15729,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-1.html#text"},{"title":"5 Experiments","offset":18200,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"Implementation Details.","offset":18218,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"Baselines and variants in Interlat.","offset":19797,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"5.1 Main Results","offset":20597,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"Latent Communication Prompts Exploration.","offset":20939,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"Semantics and Learning Dynamics.","offset":21721,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"Generalization to Symbolic Reasoning.","offset":23920,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"5.2 Compression Analysis","offset":24977,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"Compression Performance.","offset":25007,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"Why compression is effective.","offset":26381,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"5.3 Ablation Studies","offset":28476,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"5.4 Generalization to Complex Topologies.","offset":30137,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-2.html#text"},{"title":"6 Conclusion","offset":32169,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"7 Limitations","offset":33007,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"8 Ethical Considerations","offset":35460,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"9 Acknowledgments","offset":36410,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"Appendix A LLM usage","offset":36684,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"Appendix B Compression Loss","offset":37064,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"Setup.","offset":37096,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"(1) Actor cross-entropy utility.","offset":39120,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"(2) Uncertainty-weighted agreement.","offset":39880,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"(3) Latent direction alignment.","offset":42292,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"Overall objective.","offset":43651,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"B.1 Baselines and settings in Interlat.","offset":44175,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-3.html#text"},{"title":"Appendix C Benchmark","offset":49268,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"C.1 Alfworld","offset":49293,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Appendix D Ablations and Step Analysis","offset":50009,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Effect of curriculum learning.","offset":50215,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Step count versus performance.","offset":50569,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Appendix E Compression Result","offset":51830,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Appendix F Latent Parallelism Analysis","offset":52330,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Appendix G Qualitative Analysis of Latent Communication via PCA","offset":53531,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Appendix H Training Template","offset":56239,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"},{"title":"Instructions for reporting errors","offset":56756,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/part-4.html#text"}],"text_code_points":57720,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02979751e22005b3284e4d7caa19baae335ca63c9ec46e27d833e95d088e9688/source.md"}