{"id":"arxiv:2405.14838","title":"From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step","date":"2024-05-23","section":"Reasoning and the \"physics\" of language models","quality":-0.083004,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/","source_url":"https://arxiv.org/abs/2405.14838","text_available":true,"authors":["Yuntian Deng","Yejin Choi","Stuart Shieber"],"outgoing":[{"source":"arxiv:2405.14838","target":"arxiv:2206.04615","evidence":"title","id":"arxiv:2206.04615","title":"Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/70882486c1f8c46cb7b26cf100847f770eb7b39b69dee56b8a21141141ef814d/"}],"incoming":[{"source":"arxiv:2405.15071","target":"arxiv:2405.14838","evidence":"title","id":"arxiv:2405.15071","title":"Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1cd7fa83eca0b480c3bf9e2490c043b70e937777252a40b196ae7801a0620141/"},{"source":"arxiv:2412.06769","target":"arxiv:2405.14838","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:2502.05171","target":"arxiv:2405.14838","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:2509.20317","target":"arxiv:2405.14838","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:2405.14838","evidence":"id","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:2510.00184","target":"arxiv:2405.14838","evidence":"id","id":"arxiv:2510.00184","title":"Why Can't Transformers Learn Multiplication? Reverse-Engineering Reveals Long-Range Dependency Pitfalls","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7e395b2ae5bfc2a4d8bfc1e8669b4f324739b6a9b11161cba150bfdffc31746c/"},{"source":"arxiv:2609.01117","target":"arxiv:2405.14838","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/"}],"text_parts":[{"start":0,"end":15900,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html"},{"start":15900,"end":30799,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html"}],"contents":[{"title":"From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"Abstract","offset":79,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"1 Introduction","offset":1049,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"2 Background: Implicit Chain-of-Thought Reasoning","offset":4937,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"3 Stepwise Internalization","offset":7595,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"4 Experimental Setup","offset":11040,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"4.1 Data","offset":11065,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"Multi-digit multiplication.","offset":11248,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"Grade school math.","offset":11963,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"4.2 Baselines and Models","offset":12139,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"4.3 Evaluation","offset":12833,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"5 Results","offset":13500,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"Stepwise Internalization is effective.","offset":13614,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"Stepwise Internalization lags behind explicit CoT in accuracy but is faster.","offset":15399,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-1.html#text"},{"title":"6 Analysis","offset":16177,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"6.1 Accuracy-Speed Trade-off","offset":16192,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"6.2 Ablation Studies","offset":17479,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Removal smoothing.","offset":17972,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Resetting the optimizer.","offset":18682,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Removal side.","offset":19075,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Number of tokens removed per epoch.","offset":19706,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"7 Related Work","offset":20463,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"No CoT approaches.","offset":20482,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Internalizing CoT.","offset":21915,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"8 Limitations","offset":23482,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Training costs.","offset":23500,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Instability.","offset":23965,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Interpretability.","offset":24458,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Accuracy.","offset":24931,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"9 Conclusions and Future Work","offset":25821,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Acknowledgments and Disclosure of Funding","offset":27076,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Appendix A Hyperparameters","offset":27299,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Appendix B Stability Issues for Aggressive Removal","offset":28188,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Appendix C Additional Experiments","offset":28645,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Keeping position IDs.","offset":28683,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Alternative CoT formats.","offset":29112,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"},{"title":"Instructions for reporting errors","offset":29835,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/part-2.html#text"}],"text_code_points":30799,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a242b6148f6f84466ce4315ac4937278586a242dd3f847146e142f3c7e27c39c/source.md"}