{"id":"arxiv:2606.25010","title":"Emergent Capabilities Arise Randomly from Learning Sparse Attention Patterns","date":"2026-06-23","section":"Reasoning and the \"physics\" of language models","quality":0.029627,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/","source_url":"https://arxiv.org/abs/2606.25010","text_available":true,"authors":["Vatsal Baherwani","Zixi Chen","Shikai Qiu","Andrew Gordon Wilson"],"outgoing":[{"source":"arxiv:2606.25010","target":"arxiv:2001.08361","evidence":"id","id":"arxiv:2001.08361","title":"Scaling Laws for Neural Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/82302458a6d03bc051b2e4fdb8277893a81d4bd6a029969045df6a8380904eac/"},{"source":"arxiv:2606.25010","target":"arxiv:2301.05217","evidence":"id","id":"arxiv:2301.05217","title":"Progress measures for grokking via mechanistic interpretability","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f71e3b65a8b3f35beb171d17f6c65878e0914227a0042ccba59b88952554baef/"},{"source":"arxiv:2606.25010","target":"arxiv:2304.15004","evidence":"id","id":"arxiv:2304.15004","title":"Are Emergent Abilities of Large Language Models a Mirage?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d7cd8c081ad54f2fdaa2211a4b49dc21f6ff2db7282e287546d210ef90e84323/"},{"source":"arxiv:2606.25010","target":"arxiv:2405.04517","evidence":"id","id":"arxiv:2405.04517","title":"xLSTM: Extended Long Short-Term Memory","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/40fd3e03894caeb940ad2021fb4052c0d2e809706b5022f7454072997b8ba2d5/"}],"incoming":[],"text_parts":[{"start":0,"end":15967,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html"},{"start":15967,"end":31878,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html"},{"start":31878,"end":47576,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html"},{"start":47576,"end":51957,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-4.html"}],"contents":[{"title":"Emergent Capabilities Arise Randomly from Learning Sparse Attention Patterns","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"Abstract","offset":80,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"1 Introduction","offset":1468,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"2 Random Emergence of Capabilities in Language Models","offset":4957,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"2.1 Searching for Emergence During Training","offset":6557,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"2.2 Learning Core Attention Patterns Enables Abrupt Emergence","offset":9291,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"Causal Attention Head Ablation.","offset":9883,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"2.3 Emergence Across Multiple Tasks","offset":12470,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"General Trends.","offset":12900,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"3 Learning Synthetic Attention-Based Tasks","offset":14471,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"3.1 Task Definitions","offset":15036,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"Linear Map.","offset":15949,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-1.html#text"},{"title":"Cellular Automata.","offset":16917,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"3.2 Abrupt Learning with Linear Maps","offset":18838,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"3.3 Long-Context Attention Patterns are Difficult to Learn","offset":22523,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"4 The Role of Architecture in Learning Attention Patterns","offset":24998,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"4.1 Scaling Width","offset":25933,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"4.2 Scaling Attention Heads","offset":26812,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"4.3 Alternative Architectures","offset":27637,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"5 Related Work","offset":28878,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"Emergent Capabilities.","offset":28897,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"Abrupt Learning.","offset":30046,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"Neural Scaling Laws.","offset":31126,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"Our Contributions.","offset":31853,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-2.html#text"},{"title":"6 Conclusion","offset":32839,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Acknowledgements.","offset":34862,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Appendix A Emergent Capabilities in Pythia Language Models","offset":35048,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"A.1 Causal Ablations and Attention Sinks","offset":35111,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"A.2 Analyzing Emergence on Multiple Samples","offset":37221,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"A.3 Interpreting Causal Attention Mechanisms","offset":38719,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Copying.","offset":39038,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"In-Context Repetition.","offset":39458,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Pattern Completion.","offset":40116,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Indirect Object Identification.","offset":40731,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Appendix B Abrupt Learning on Synthetic Tasks","offset":41365,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"B.1 Notation","offset":41415,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"B.2 Attention Entropy","offset":41537,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"B.3 Attention Interventions","offset":42019,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"B.4 What Variables Affect Loss Plateau Length?","offset":42542,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"B.5 Depth Scaling","offset":44807,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"B.6 Comparing Alternative Attention Mechanisms","offset":46304,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Appendix C Does Pre-pretraining Accelerate Emergence of Capabilities?","offset":47502,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-3.html#text"},{"title":"Hyperparameters.","offset":49054,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-4.html#text"},{"title":"Instructions for reporting errors","offset":50993,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/part-4.html#text"}],"text_code_points":51957,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f87db387bad54ae4fbc0589d2a702f78aeb0817b28f20844f6502bc552a89c2f/source.md"}