{"id":"arxiv:2606.31779","title":"Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers","date":"2026-06-30","section":"Reasoning and the \"physics\" of language models","quality":0.291912,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/","source_url":"https://arxiv.org/abs/2606.31779","text_available":true,"authors":["Ying Fan","Anej Svete","Kangwook Lee"],"outgoing":[{"source":"arxiv:2606.31779","target":"arxiv:1807.03819","evidence":"id","id":"arxiv:1807.03819","title":"Universal Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1d843516d980c3404cb622a2a490edb535a3afe27a85a980335a54b5fcd2220a/"},{"source":"arxiv:2606.31779","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:2606.31779","target":"arxiv:2412.06769","evidence":"title","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:2606.31779","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:2606.31779","target":"arxiv:2506.08007","evidence":"title","id":"arxiv:2506.08007","title":"Reinforcement Pre-Training","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a5ae77d4a2dcd6843e856083ecc6c30a72563e6896503d6efdfae117b8b007e6/"},{"source":"arxiv:2606.31779","target":"arxiv:2507.10524","evidence":"title","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:2606.31779","target":"arxiv:2509.20317","evidence":"id","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:2606.31779","target":"arxiv:2509.25239","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:2606.31779","target":"arxiv:2605.19376","evidence":"id","id":"arxiv:2605.19376","title":"Generative Recursive Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/795bad9bb1b1a063090fbb925ba6e68bf30211de0ecffd75d56f858b28321ae7/"}],"incoming":[],"text_parts":[{"start":0,"end":15682,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html"},{"start":15682,"end":31384,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html"},{"start":31384,"end":46850,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html"},{"start":46850,"end":61923,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html"},{"start":61923,"end":77395,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html"},{"start":77395,"end":82279,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-6.html"}],"contents":[{"title":"Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"Abstract","offset":83,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"1 Introduction","offset":1708,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"2 Preliminaries","offset":8381,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"Explicit CoT.","offset":8469,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"Latent reasoning","offset":9078,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"Looped padded Transformers.","offset":9473,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"Prior latent reasoning methods.","offset":9917,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"3 Method","offset":11103,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"3.1 Padded Latent Prefix","offset":12154,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"3.2 Looped Latent Computation","offset":13836,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"3.3 Direct Step-Aligned Supervision","offset":15585,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"3.3.1 Objective","offset":15626,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"Step CoT supervision loss.","offset":15648,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-1.html#text"},{"title":"Answer supervision loss.","offset":16676,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Key properties of the objective.","offset":18211,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Inference.","offset":18735,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"No latent decoding for answer generation.","offset":19337,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"3.3.2 Roles of ℒ step {\\mathcal{L}_{\\mathrm{step}}} and ℒ ans {\\mathcal{L}_{\\mathrm{ans}}}","offset":19786,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Parallel chain likelihood.","offset":20316,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Complementary roles.","offset":21323,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"3.4 LOTUS-aux: An Auxiliary Decoder Variant","offset":23154,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Latent supervision via the auxiliary decoder.","offset":24289,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Per-iteration supervision and the full objective.","offset":26752,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Inference.","offset":27540,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"4 Experiments","offset":27830,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"4.1 Setup","offset":28484,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Datasets and backbones.","offset":28499,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"Methods.","offset":29167,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"4.2 Main Results","offset":30416,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"LOTUS matches explicit CoT at scale, where prior latent methods fall behind.","offset":30438,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"The two routings match at 3B, but only direct LOTUS stays robust at smaller scales.","offset":31293,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-2.html#text"},{"title":"LOTUS reasons 2.5 × 2.5\\times faster than explicit CoT.","offset":31793,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Speedup grows to 6.9 × 6.9\\times on natural-language CoT traces.","offset":33341,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"4.3 Ablation Studies","offset":34096,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"4.3.1 Latent Supervision Design","offset":34294,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Both the looped backbone and latent supervision are beneficial.","offset":35299,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"The best supervision schedule differs between LOTUS and LOTUS-aux.","offset":35966,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"4.3.2 Looped Architecture Design","offset":36404,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Latent token budget per block c c .","offset":36443,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Inference-time c c .","offset":36964,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Looped iterations R {R} .","offset":37719,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Inference-time R {R} .","offset":38286,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"5 Latent Representation Analysis","offset":38807,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"5.1 Likelihood of the Gold Chain under Different Readout Options","offset":39564,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Metrics.","offset":40958,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"LOTUS’s post-loop latent 𝒉 ( R ) {{{\\bm{h}}}^{({R})}} contains the gold CoT via a direct readout.","offset":41752,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"LOTUS-aux has CoT signal in both 𝒉 ( R ) {{{\\bm{h}}}^{({R})}} and 𝒛 ~ ( t ) {\\widetilde{{{\\bm{z}}}}^{(t)}} .","offset":42415,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"5.2 Mass on Unseen Valid Chains","offset":43350,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Setup.","offset":43635,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Findings.","offset":44300,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"5.3 Disentangling the Two Losses in LOTUS","offset":45210,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Gold-chain NLL.","offset":45669,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"ℒ step {\\mathcal{L}_{\\mathrm{step}}} and ℒ ans {\\mathcal{L}_{\\mathrm{ans}}} play complementary roles.","offset":45978,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"6 Conclusion","offset":46833,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-3.html#text"},{"title":"Limitations.","offset":47690,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Acknowledgments","offset":48260,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Appendix A Related Work","offset":48354,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Looped and recurrent-depth transformers.","offset":48687,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Diffusion language models.","offset":51668,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Pause and padding symbols.","offset":52450,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Compressing and replacing discrete CoT.","offset":53107,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Speculative decoding.","offset":55623,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Appendix B Notation","offset":56125,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Appendix C Closely Related Latent Reasoning Methods","offset":56210,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Coconut [ Hao et al., 2025 ]","offset":56388,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"CODI [ Shen et al., 2025 ] and SIM-CoT [ Wei et al., 2025 ]","offset":57110,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Parallel Continuous CoT (PCCoT) [ Wu et al., 2025 ] and KaVa [ Kuzina et al., 2026 ]","offset":57841,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Comparison with latent methods at other compute budgets.","offset":59132,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Natural-language CoT stress-test.","offset":60573,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Appendix D Experimental Details","offset":61358,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Latent prefix budget.","offset":61471,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Training curriculum.","offset":61895,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-4.html#text"},{"title":"Optimization details.","offset":62899,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Natural-language training setup.","offset":63653,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"LOTUS-aux auxiliary decoder.","offset":64115,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Evaluation.","offset":64519,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Batched KV-cache sharing.","offset":64817,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Appendix E PCL Versus Autoregressive NLL Lower Bounds","offset":65087,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Appendix F Details for Latent Representation Analysis ( Section 5 )","offset":66227,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"F.1 Greedy Readout Examples","offset":66299,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Example 1 — Janet’s ducks.","offset":67782,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Example 2 — James’s sprints.","offset":69789,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Example 3 — Carla’s download (failure).","offset":71655,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Example 4 — Melanie’s vacuums (failure).","offset":73740,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"F.2 Details for Multi-Path Analysis ( Section 5.2 )","offset":76524,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Datasets.","offset":76581,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Path bank.","offset":76933,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Candidate sets and filtering.","offset":77358,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-5.html#text"},{"title":"Metrics.","offset":78388,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-6.html#text"},{"title":"F.3 Multi-Path Readout Example","offset":79589,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-6.html#text"},{"title":"Instructions for reporting errors","offset":81315,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/part-6.html#text"}],"text_code_points":82279,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bf07d5c19e507d9ec91a48f9067e6c29fbe8965d001923e2de0e96da143c72e1/source.md"}