{"id":"arxiv:2501.14082","title":"Communicating Activations Between Language Model Agents","date":"2025-01-23","section":"LLMs: architectures, context, training","quality":0.117038,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/","source_url":"https://arxiv.org/abs/2501.14082","text_available":true,"authors":["Vignav Ramesh","Kenneth Li"],"outgoing":[{"source":"arxiv:2501.14082","target":"arxiv:2203.11171","evidence":"id","id":"arxiv:2203.11171","title":"Self-Consistency Improves Chain of Thought Reasoning in Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02a15b098d96a94d463bb470f59a34723abc34e59196f35c539f9796d8f3c3df/"},{"source":"arxiv:2501.14082","target":"arxiv:2405.07987","evidence":"id","id":"arxiv:2405.07987","title":"The Platonic Representation Hypothesis","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6c7695929f8bbcea43ee98da9cd8f2fe806f2ba292f0b4e3e7fea2b64254afb4/"},{"source":"arxiv:2501.14082","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:2501.14082","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:2501.14082","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:2501.14082","target":"url:openai.com/research/weak-to-strong-generalization","evidence":"title","id":"url:openai.com/research/weak-to-strong-generalization","title":"Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fb7b23a617dce974dbb70db824c97c1de84afa11e1e290d5cb9f5bd88857e762/"}],"incoming":[{"source":"arxiv:2511.09149","target":"arxiv:2501.14082","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:2608.11676","target":"arxiv:2501.14082","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":14463,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html"},{"start":14463,"end":29698,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html"},{"start":29698,"end":45158,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html"}],"contents":[{"title":"Communicating Activations Between Language Model Agents","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"Abstract","offset":59,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"Keywords:","offset":1518,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"1 Introduction","offset":1536,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"2 Related Work","offset":6636,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"Multi-agent communication","offset":6655,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"Activation engineering","offset":9862,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"Model composition and grafting","offset":10607,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"3 Communicating Activations Between Language Models","offset":12511,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"3.1 Method","offset":12870,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-1.html#text"},{"title":"3.2 Compute Analysis","offset":17340,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"3.3 Why should this work?","offset":19498,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"4 Experiments","offset":22001,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"4.1 Multi-player coordination games","offset":22216,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"4.2 Reasoning Benchmarks","offset":24582,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"Baselines","offset":24724,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"Models","offset":26103,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"Datasets","offset":27393,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"Metrics","offset":28668,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"Comprehensive evaluation with the 𝙻𝙻𝚊𝙼𝙰 \\mathtt{LLaMA} family","offset":29630,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-2.html#text"},{"title":"Performance-compute tradeoff and generalization to different model scales","offset":31588,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"Communication across model families","offset":32726,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"5 Conclusion","offset":33885,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"Impact Statement","offset":36591,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"Acknowledgements","offset":36812,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"Appendix A Qualitative Results","offset":36948,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"Appendix B Additional Experiments","offset":36983,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"B.1 Modifying Activations of All Tokens","offset":37021,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"B.2 Incorporating Chain-of-Thought Prompting","offset":38861,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"B.3 Learning 𝑾 \\bm{W} In-Distribution","offset":39719,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"B.4 Activation Space Similarity ∝ \\propto AC Performance Gain","offset":40981,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"B.5 Communicating Activations Between Identical Models","offset":42074,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"B.6 Additional Rounds of Natural Language Debate","offset":43358,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"B.7 Full MMLU Benchmark Results","offset":43874,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"},{"title":"Instructions for reporting errors","offset":44194,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/part-3.html#text"}],"text_code_points":45158,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e67d4addcda44093e242074332925d000726f87bd9538f6b318bdfc454a7dc13/source.md"}