{"id":"arxiv:2608.09888","title":"BDH-CQ: In-Context Learning with Recurrent Latent Reasoning","date":"2026-08-10","section":"Post-training","quality":-0.339191,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/","source_url":"https://arxiv.org/abs/2608.09888","text_available":true,"authors":["Björn Engdahl","Adrian Kosowski","Jan Chorowski"],"outgoing":[{"source":"arxiv:2608.09888","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:2608.09888","target":"arxiv:2412.06769","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:2608.09888","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:2608.09888","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:2608.09888","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:2608.09888","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/"}],"incoming":[],"text_parts":[{"start":0,"end":15992,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html"},{"start":15992,"end":31907,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html"},{"start":31907,"end":45326,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html"}],"contents":[{"title":"BDH-CQ: In-Context Learning with Recurrent Latent Reasoning","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"Abstract","offset":63,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"1 Introduction","offset":956,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"2 ARC as a controlled, verifiable testbed for BDH-CQ’s in-context generalization","offset":5575,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"3 Introducing BDH-CQ: In-context learning through recurrent memory and latent reasoning","offset":7219,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"3.1 BDH provenance","offset":7311,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"3.2 In-context learning through recurrent memory","offset":8661,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"3.3 Recurrent latent reasoning","offset":9877,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"4 Training data and objective","offset":11016,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"4.1 ARC task formulation","offset":11050,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"4.2 Training mixture","offset":11747,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"5 Public ARC-AGI-1 evaluation","offset":12199,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"Independent evaluation.","offset":13685,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"6 Behavioral analysis on ARC-like tasks","offset":14127,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"6.1 Concept-organized capability profile","offset":14414,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"6.2 Simple operators extrapolate; ordering and nesting expose distinct boundaries","offset":15905,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-1.html#text"},{"title":"6.3 Context binds dense mappings, but composition is operation-dependent","offset":19271,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"6.4 Within-task consistency","offset":21717,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"6.5 Identifier and batch-context replication","offset":22335,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"6.6 Attempt delivery, repeatability, and effort","offset":23607,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"7 Scaling reasoning effort increases pass@2","offset":24610,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"8 Related work","offset":25263,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"Verbalized reasoning.","offset":25282,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"Continuous thoughts in language models.","offset":25721,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"Compressed discrete and communicated latent states.","offset":27270,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"Recurrent depth and looped models.","offset":28036,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"Task-trained recursive solvers.","offset":28803,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"ARC and concept-organized evaluation.","offset":30008,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"9 Discussion","offset":30390,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-2.html#text"},{"title":"9.1 Outlook","offset":32206,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"9.2 Scaling model size","offset":34421,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"10 Conclusion","offset":34827,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"Appendix A Evaluation set analysis","offset":35684,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"A.1 Evaluation sets and surface descriptors","offset":35723,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"A.2 Transformation mechanics","offset":36763,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"A.3 Controlled structural manipulations","offset":38851,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"A.4 Failure structure and scope","offset":42130,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"},{"title":"Instructions for reporting errors","offset":44362,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/part-3.html#text"}],"text_code_points":45326,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f0c77a8a0f50254e9efdf9a511e08fc2a40677bfa029864bdd0a9cbce561c666/source.md"}