{"id":"arxiv:2608.17981","title":"Recirculation","date":"2026-08-18","section":"LLMs: architectures, context, training","quality":null,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/","source_url":"https://arxiv.org/abs/2608.17981","text_available":true,"authors":["Michael C. Mozer","Shoaib Ahmed Siddiqui","Danny Sawyer","Sunny Sanyal"],"outgoing":[{"source":"arxiv:2608.17981","target":"arxiv:1807.03819","evidence":"title","id":"arxiv:1807.03819","title":"Universal Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1d843516d980c3404cb622a2a490edb535a3afe27a85a980335a54b5fcd2220a/"},{"source":"arxiv:2608.17981","target":"arxiv:2006.16236","evidence":"title","id":"arxiv:2006.16236","title":"Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d119844a2dce8a3b156558743b90add84774d00894634b6478d03abb4f5216f9/"},{"source":"arxiv:2608.17981","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:2608.17981","target":"arxiv:2502.05171","evidence":"title","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:2608.17981","target":"arxiv:2504.13837","evidence":"title","id":"arxiv:2504.13837","title":"Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b90f70269aeb02311d74ffaadf47e139683424a47118dc18619000ba72174d17/"},{"source":"arxiv:2608.17981","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:2608.17981","target":"arxiv:2510.04871","evidence":"id","id":"arxiv:2510.04871","title":"Less is More: Recursive Reasoning with Tiny Networks","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4345eebe9dc7e67f82617896772f55f0970971516995c1688b6812f9ebf054f6/"},{"source":"arxiv:2608.17981","target":"arxiv:2604.17121","evidence":"id","id":"arxiv:2604.17121","title":"The Topological Trouble With Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7f9d762bbe192e1208f6d7dc289b7e79c23d4bdde0b1b3999ff5fda55b0dcd1a/"}],"incoming":[{"source":"arxiv:2604.17121","target":"arxiv:2608.17981","evidence":"id","id":"arxiv:2604.17121","title":"The Topological Trouble With Transformers","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/7f9d762bbe192e1208f6d7dc289b7e79c23d4bdde0b1b3999ff5fda55b0dcd1a/"}],"text_parts":[{"start":0,"end":15187,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-1.html"},{"start":15187,"end":30095,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html"},{"start":30095,"end":46034,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html"},{"start":46034,"end":61757,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html"},{"start":61757,"end":69380,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-5.html"}],"contents":[{"title":"Recirculation","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-1.html#text"},{"title":"Abstract","offset":17,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-1.html#text"},{"title":"1 State tracking","offset":1556,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-1.html#text"},{"title":"2 Proposed method: Recirculation","offset":9419,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-1.html#text"},{"title":"3 Related research","offset":15164,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-1.html#text"},{"title":"4 Experiments","offset":21083,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.1 Hyperparameter sweeps","offset":21101,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.2 Perplexity evaluation","offset":23229,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.3 Normalization and ramping","offset":24525,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.4 Analysis of recirculation","offset":26150,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.4.1 Robustness across model families","offset":26185,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.4.2 Recirculation versus temperature tuning","offset":27852,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.4.3 Recirculation versus looping","offset":28769,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.4.4 Which tokens benefit from recirculation?","offset":30042,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-2.html#text"},{"title":"4.5 Generative tasks","offset":32895,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html#text"},{"title":"4.5.1 Instruction following","offset":33130,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html#text"},{"title":"4.5.2 Contextualization","offset":34191,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html#text"},{"title":"4.5.3 Multiple-choice and single-token response tasks","offset":36079,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html#text"},{"title":"4.5.4 GSM8k","offset":37296,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html#text"},{"title":"4.6 Adaptive recirculation","offset":39372,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html#text"},{"title":"5 Discussion","offset":45209,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-3.html#text"},{"title":"5.1 Model-design affordances","offset":46806,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"5.2 Limitations and future directions","offset":48528,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"6 Acknowledgements","offset":52997,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"Appendix A Unrolled recirculation architecture","offset":53335,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"Appendix B Recirculation implementation details","offset":53582,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"B.1 Hyperparameter sweeps","offset":53634,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"B.2 Perplexity evaluation","offset":54164,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"B.3 Normalization and ramping","offset":55111,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"Appendix C Basic results","offset":57458,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"C.1 Robustness across architectures","offset":57487,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"C.2 Recirculation versus temperature tuning","offset":59511,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"C.3 Recirculation versus looping","offset":60283,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"C.4 Which tokens benefit from recirculation?","offset":60819,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"Appendix D Downstream evaluations","offset":61637,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"D.1 Instruction following","offset":61675,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-4.html#text"},{"title":"D.2 Contextualization","offset":63179,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-5.html#text"},{"title":"D.3 Multiple-choice and single-token response tasks","offset":64006,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-5.html#text"},{"title":"D.4 Standard benchmark datasets: GSM8k","offset":64658,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-5.html#text"},{"title":"D.5 Adaptive recirculation","offset":65097,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-5.html#text"},{"title":"Instructions for reporting errors","offset":68416,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/part-5.html#text"}],"text_code_points":69380,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/fd91ffc957070136795d3728f662c997738234065b624eb720acd8547d2edaf6/source.md"}