{"id":"arxiv:2412.06769","title":"Training Large Language Models to Reason in a Continuous Latent Space","date":"2024-12-09","section":"Reasoning and the \"physics\" of language models","quality":0.149043,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/","source_url":"https://arxiv.org/abs/2412.06769","text_available":true,"authors":["Shibo Hao","Sainbayar Sukhbaatar","DiJia Su","Xian Li"],"outgoing":[{"source":"arxiv:2412.06769","target":"arxiv:2402.03300","evidence":"id","id":"arxiv:2402.03300","title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/636fb871904bc06f742cfdea069796f9e9665f95bd58dc63260eaff5154ef985/"},{"source":"arxiv:2412.06769","target":"arxiv:2405.14838","evidence":"id","id":"arxiv:2405.14838","title":"From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/"},{"source":"arxiv:2412.06769","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:2412.06769","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:2412.06769","target":"doi:10.1038/s41586-024-07522-w","evidence":"title","id":"doi:10.1038/s41586-024-07522-w","title":"Language is primarily a tool for communication rather than thought","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60cfed62cd01166a972691a7c0da286223d861d23feebc5fadd3fa446168a33c/"},{"source":"arxiv:2412.06769","target":"url:ai.meta.com/research/publications/large-concept-models-language-modeling-in-a-sentence-representation-space","evidence":"id","id":"url:ai.meta.com/research/publications/large-concept-models-language-modeling-in-a-sentence-representation-space","title":"Large Concept Models: Language Modeling in a Sentence Representation Space","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a5cd322d52f1966361afb1d5225560e2d3ce4d22a117bd14cfa27ec669e09673/"}],"incoming":[{"source":"arxiv:2501.14082","target":"arxiv:2412.06769","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:2502.05171","target":"arxiv:2412.06769","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:2506.21734","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2506.21734","title":"Hierarchical Reasoning Model","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3604312101ff09ab9903e7d6a32c7b2b8a9a7ece6a9d145d04ea01c9d13ffa3a/"},{"source":"arxiv:2507.10524","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2507.10524","title":"Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7b25702c35f07446cea97d6a98afc489c8fdda54560c05fbeb1b3b139a9da622/"},{"source":"arxiv:2509.20317","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2509.20317","title":"SIM-CoT: Supervised Implicit Chain-of-Thought","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/72b64dd6a30c71d99920cafabbb2c909d03febab8991eeaf3c091b16cb7ed348/"},{"source":"arxiv:2509.25239","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2509.25239","title":"A Formal Comparison Between Chain of Thought and Latent Thought","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8399816be05fff6a7002f5c3439ec43ffe7b468424759b4cc70f5efc7739ea9e/"},{"source":"arxiv:2511.09149","target":"arxiv:2412.06769","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:2601.21343","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2601.21343","title":"Self-Improving Pretraining: using post-trained models to pretrain better models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/43ce1b189d6b21d445c0ce6d4d04b4352d268beb4346da43be19982402befde8/"},{"source":"arxiv:2604.17121","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2604.17121","title":"The Topological Trouble With Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7f9d762bbe192e1208f6d7dc289b7e79c23d4bdde0b1b3999ff5fda55b0dcd1a/"},{"source":"arxiv:2605.12466","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2605.12466","title":"Solve the Loop: Attractor Models for Language and Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/eb46f81cd5718eb6b9256c583cf9dd2e97b793b4d14337e487c16285f96e2b9a/"},{"source":"arxiv:2605.19376","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2605.19376","title":"Generative Recursive Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/795bad9bb1b1a063090fbb925ba6e68bf30211de0ecffd75d56f858b28321ae7/"},{"source":"arxiv:2605.21488","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2605.21488","title":"Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48b3beca0f016a3683df4c6c71fc4be9f13a2c12c61d4ad17d1117c6139bb11e/"},{"source":"arxiv:2606.18206","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2606.18206","title":"Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/37ea2d83a92ece4ab3ba4434af71f291d1fdc5040676f75110210f40aad3b81a/"},{"source":"arxiv:2606.31779","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2606.31779","title":"Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/"},{"source":"arxiv:2608.09888","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2608.09888","title":"BDH-CQ: In-Context Learning with Recurrent Latent Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/"},{"source":"arxiv:2608.13040","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2608.13040","title":"Latent On-Policy Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/"},{"source":"arxiv:2608.17981","target":"arxiv:2412.06769","evidence":"title","id":"arxiv:2608.17981","title":"Recirculation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/"},{"source":"arxiv:2609.01117","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2609.01117","title":"Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e9d4f61afc3930fa770d5ea53d3f43a5157fa4904fcdfaa514786e617905e20b/"},{"source":"url:ai-2027.com","target":"arxiv:2412.06769","evidence":"id","id":"url:ai-2027.com","title":"AI 2027 (Kokotajlo)","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3fd447e0ff58689ae00a4666e28bbd7faecd472dec7195ab1d524cb5402a1766/"}],"text_parts":[{"start":0,"end":15912,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-1.html"},{"start":15912,"end":31499,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html"},{"start":31499,"end":40289,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html"}],"contents":[{"title":"Training Large Language Models to Reason in a Continuous Latent Space","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-1.html#text"},{"title":"Abstract","offset":73,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-1.html#text"},{"title":"1 Introduction","offset":1489,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-1.html#text"},{"title":"2 Related Work","offset":5673,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-1.html#text"},{"title":"3 Coconut : Chain of Continuous Thought","offset":10296,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-1.html#text"},{"title":"4 Continuous Space Enables Latent Tree Search","offset":15862,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-1.html#text"},{"title":"4.1 Experimental Setup","offset":16948,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"4.2 Overall Results","offset":20698,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"4.3 Interpreting the Latent Reasoning as Tree Search","offset":22095,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"4.4 Why is Latent Space Better for Planning?","offset":24618,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"5 Empirical Results with Coconut","offset":26285,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"5.1 Experimental Setup","offset":26655,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"5.2 Baselines and Variants of Coconut","offset":28344,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"5.3 Results and Discussion","offset":29780,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-2.html#text"},{"title":"6 Conclusion","offset":33836,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"Appendix A Datasets","offset":34841,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"A.1 Examples","offset":34865,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"A.2 Construction of ProsQA","offset":34986,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"A.3 Statistics","offset":36559,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"Appendix B Clock-Time Reasoning Efficiency Metric","offset":36626,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"Appendix C More Discussion","offset":37170,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"C.1 Using More Continuous Thoughts","offset":37201,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"C.2 Coconut with Larger Models","offset":38060,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"},{"title":"Instructions for reporting errors","offset":39325,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/part-3.html#text"}],"text_code_points":40289,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/source.md"}