{"id":"arxiv:2607.27372","title":"Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation","date":"2026-07-29","section":"Data, training, optimization","quality":0.296593,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/","source_url":"https://arxiv.org/abs/2607.27372","text_available":true,"authors":["Alexi Gladstone","Heng Ji","Yilun Du"],"outgoing":[{"source":"arxiv:2607.27372","target":"arxiv:2301.08243","evidence":"title","id":"arxiv:2301.08243","title":"Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/475bac0a3b29edd42642bbe10d1a41bfb18a3022ac95b9b1a3db9e5d5d0b95fe/"},{"source":"arxiv:2607.27372","target":"arxiv:2304.07193","evidence":"title","id":"arxiv:2304.07193","title":"DINOv2: Learning Robust Visual Features without Supervision","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b15933596df35497024556ae88bfa696f3b0e8372b678194f04f1bfb9f3be553/"},{"source":"arxiv:2607.27372","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/"}],"incoming":[],"text_parts":[{"start":0,"end":14794,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-1.html"},{"start":14794,"end":30792,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html"},{"start":30792,"end":46260,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html"},{"start":46260,"end":62048,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html"},{"start":62048,"end":77719,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html"},{"start":77719,"end":93625,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html"},{"start":93625,"end":109461,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html"},{"start":109461,"end":114637,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html"}],"contents":[{"title":"Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-1.html#text"},{"title":"Abstract","offset":86,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-1.html#text"},{"title":"1 Introduction","offset":2570,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-1.html#text"},{"title":"2 Background","offset":9135,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-1.html#text"},{"title":"2.1 Mode Forcing","offset":9668,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-1.html#text"},{"title":"2.2 End-to-End Generation","offset":16277,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"2.3 Existing Generative Models","offset":18404,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"3 Explorative Modeling Approach","offset":19550,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"3.1 Explorative Modeling Intuition","offset":19586,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"Scalable Training of Latent Variable Models","offset":21993,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"A Scalable Way to Resolve Coupling","offset":23012,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"Generative Modeling via Search","offset":23884,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"Spreading Predictions Across Modes","offset":24233,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"Minimizing an Implicit Energy","offset":24914,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"3.2 Forward and Reverse Explorative Modeling","offset":25351,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"Forward XM.","offset":25528,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"Reverse XM.","offset":26124,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"What Forward and Reverse XM Optimize.","offset":27552,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"4 Experimentation and Results","offset":30758,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-2.html#text"},{"title":"4.1 Explorative Modeling as a New Scaling Axis","offset":32429,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Does Exploration Improve Existing Generative Models’ Performance?","offset":32481,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Does Performance Across Modalities Scale with Exploration?","offset":35673,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Does Exploration Improve Generalization?","offset":36868,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Does Exploration Improve State-of-the-Art Recipes at Scale?","offset":38768,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Do Improvements From Exploration Vary With Scale?","offset":39908,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Which Models Benefit Most From Exploration?","offset":42057,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"4.2 End-to-End Explorative Models","offset":44207,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Can Explorative Modeling be Used for End-to-End Generation?","offset":44246,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"5 Discussion","offset":46200,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"Mode Forcing as a Predictive Theory.","offset":46217,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-3.html#text"},{"title":"The Benefits of Surplus Generative Expressivity.","offset":48370,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"6 Future Works and Broader Impact","offset":49657,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Exploration as a Scaling Axis for More Generative Models.","offset":49779,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"End-to-end XM Applications.","offset":51405,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Improving and Understanding XMs.","offset":52657,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Scaling Reverse XMs.","offset":54702,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Exploration beyond Pretraining.","offset":56089,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"7 Limitations and Conclusion","offset":56666,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Limitations.","offset":57025,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Conclusion.","offset":60273,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Author Contributions","offset":61600,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Alexi Gladstone","offset":61625,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Heng Ji and Yilun Du","offset":61821,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Acknowledgement","offset":62028,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-4.html#text"},{"title":"Appendix A Additional Experimentation","offset":63944,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"A.1 Exact Results for Exploration Scaling","offset":64582,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"A.2 Comparing Exploration to Minibatch Optimal Transport Couplings","offset":64755,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"A.3 Reverse XM Language Models","offset":65881,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Appendix B Additional Intuition","offset":67028,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Generative Modeling with a For Loop.","offset":67064,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Exploration as Building an Associative Memory.","offset":67545,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Appendix C Approach Details","offset":68466,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Batching Forward XM.","offset":68498,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Other Ways to Explore.","offset":68941,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Gradients and Memory.","offset":69555,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Exploration FLOP Cost.","offset":70227,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Generating the K K Candidates.","offset":71049,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Implementing Reverse XM.","offset":71695,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Appendix D Experimental Details","offset":72486,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Model Sizes.","offset":72684,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Image Generation.","offset":72838,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"RAE Image Generation.","offset":74061,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Video Generation.","offset":74814,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Individual Results in More Detail.","offset":75465,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Behavior Cloning.","offset":76047,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Goal-Conditioned World Modeling.","offset":76434,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Language Modeling.","offset":76986,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"Appendix E Related Works","offset":77650,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"E.1 End-to-End Generative Modeling","offset":77679,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-5.html#text"},{"title":"E.2 Coupling","offset":79913,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"E.3 Explorative Modeling Based Methods","offset":80780,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"Appendix F Additional Theory","offset":84681,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"F.1 Generative Expressivity Details","offset":84714,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"Scope of E E .","offset":86989,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"Per-Prediction Expressivity.","offset":88287,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"F.2 What Forward and Reverse XM Optimize","offset":92113,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"Overview.","offset":92159,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"Setup.","offset":93612,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-6.html#text"},{"title":"Proposition 1 (Forward XM is maximum likelihood).","offset":94828,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Proof sketch.","offset":95687,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Remark (the optimum at fixed σ \\sigma ).","offset":96477,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"The role of K K .","offset":96818,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"The hard min.","offset":98108,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Proposition 2 (Reverse XM targets the reverse KL, but collapses without an entropy term).","offset":98720,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Proof sketch.","offset":100091,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Where the blur sits.","offset":101029,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"On the ELBO.","offset":101344,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Assumptions.","offset":102035,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Proposition 3 (Expressivity of smooth Forward XM under separation).","offset":103862,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Proof sketch.","offset":105186,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Appendix G Frequently Asked Questions","offset":107731,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"What’s the main takeaway? How is this paper ‘novel’ if it’s just best-of- K K ?","offset":107839,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Does exploration make inference more expensive?","offset":109407,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-7.html#text"},{"title":"Is exploration really a new “axis” if it just costs more compute?","offset":109734,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html#text"},{"title":"Why the name Explorative Modeling?","offset":110346,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html#text"},{"title":"When does exploration help?","offset":111076,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html#text"},{"title":"Does training on the model’s own generations cause collapse?","offset":111501,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html#text"},{"title":"How should the amount of exploration K K be chosen?","offset":112116,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html#text"},{"title":"Is Explorative Modeling a form of reinforcement learning?","offset":113244,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html#text"},{"title":"Instructions for reporting errors","offset":113673,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/part-8.html#text"}],"text_code_points":114637,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/542a5e85b84624381953f643ec6075a588c08c542c4f945e57d39c9164c9a835/source.md"}