{"id":"arxiv:2604.11791","title":"A Mechanistic Analysis of Looped Reasoning Language Models","date":"2026-04-13","section":"Reasoning and the \"physics\" of language models","quality":-0.159909,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/","source_url":"https://arxiv.org/abs/2604.11791","text_available":true,"authors":["Hugh Blayney","Álvaro Arroyo","Johan Obando-Ceron","Pablo Samuel Castro"],"outgoing":[{"source":"arxiv:2604.11791","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:2604.11791","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:2604.11791","target":"arxiv:2506.21734","evidence":"id","id":"arxiv:2506.21734","title":"Hierarchical Reasoning Model","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3604312101ff09ab9903e7d6a32c7b2b8a9a7ece6a9d145d04ea01c9d13ffa3a/"},{"source":"arxiv:2604.11791","target":"arxiv:2507.10524","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:2604.11791","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:2604.11791","target":"url:pub.sakana.ai/ctm","evidence":"id","id":"url:pub.sakana.ai/ctm","title":"Continuous Thought Machines","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2d6fdb072de0d3edb79a12bb935f300e9fe9b517f681e39459a39db71cfd44f8/"}],"incoming":[{"source":"arxiv:2605.12466","target":"arxiv:2604.11791","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:2605.21488","target":"arxiv:2604.11791","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:2604.11791","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/"}],"text_parts":[{"start":0,"end":15342,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html"},{"start":15342,"end":31256,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html"},{"start":31256,"end":47111,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html"},{"start":47111,"end":62819,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html"}],"contents":[{"title":"A Mechanistic Analysis of Looped Reasoning Language Models","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"Abstract","offset":62,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"Keywords:","offset":1486,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"1 Introduction","offset":1538,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"2 Preliminaries","offset":3641,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"2.1 Looped Transformers","offset":3661,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"2.2 Stages of Inference","offset":10460,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"3 Related Work","offset":12370,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"Looped and Recurrent Transformers","offset":12389,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"Stages of Inference and Attention Dynamics","offset":14497,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"Test-time Computation","offset":15314,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-1.html#text"},{"title":"4 Looped Transformers Tend Towards the Same Attention Patterns","offset":16507,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"Proposition 4.1 (Cyclic recurrent blocks reach cyclic fixed points) .","offset":17274,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"Proposition 4.2 (Recurrent attention patterns change slowly under state convergence) .","offset":18568,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"4.1 Empirical Validation","offset":20056,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"4.2 Impact of Architecture Choices","offset":22823,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"5 Stages of Inference in Looped Models Mirror Feedforward Computation","offset":24469,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"5.1 Self-Organization Into Stages of Inference","offset":27602,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"5.2 Stability to Unseen Numbers of Recurrences","offset":29859,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-2.html#text"},{"title":"6 Conclusion","offset":31715,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Implications of Findings","offset":32288,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Limitations and Future Work","offset":33496,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Impact Statement","offset":34095,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Acknowledgments","offset":34699,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Appendix Contents","offset":34996,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Appendix A Proofs of Propositions","offset":35018,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Proof of Proposition 4.1","offset":35056,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Proof sketch.","offset":35087,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Proof.","offset":35864,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Proof of Proposition 4.2","offset":38793,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Proof.","offset":38824,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Appendix B Additional Experimental Details","offset":43007,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Appendix C Non-Fixed-Point Limiting Behavior","offset":44753,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"C.1 How Frequent is Non-Fixed-Point Behavior?","offset":44802,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Long Persona","offset":45838,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Long Persona (Padded)","offset":45944,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"Short Math","offset":46150,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-3.html#text"},{"title":"C.2 Do Intermediate Layers Exhibit Non-Fixed-Point Behavior?","offset":47755,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"C.3 How Does Non-Fixed-Point Behavior Impact Stages of Inference?","offset":48993,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"Appendix D Additional Fixed Point Results","offset":49988,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"D.1 Cyclic Similarity","offset":50034,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"D.2 Fixed Point and Successive Differences","offset":50821,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"D.3 Latent Space Trajectories","offset":51698,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"D.4 Architecture Choices","offset":52853,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"Appendix E Additional Stages of Inference Results","offset":53256,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"E.1 Input Independent Metrics","offset":53466,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"E.2 Input Dependent Metrics","offset":54620,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"E.3 Cyclic Stages of Inference","offset":56662,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"E.4 Non-Reasoning Stages of Inference","offset":59158,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"E.5 Stability To Unseen Test-Time Recurrences","offset":60100,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"E.6 How Architecture Choices Affect the Formation of Stages of Inference","offset":60633,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"Appendix F Looped Floorplan","offset":61425,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"},{"title":"Instructions for reporting errors","offset":61855,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/part-4.html#text"}],"text_code_points":62819,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ae8c682d5708c6c47d65c4eb540dd5bccb877c5e332852765507128d6d1b985d/source.md"}