{"id":"arxiv:2602.06130","title":"Self-Improving World Modelling with Latent Actions","date":"2026-02-05","section":"Reinforcement learning","quality":null,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/","source_url":"https://arxiv.org/abs/2602.06130","text_available":true,"authors":["Yifu Qiu","Zheng Zhao","Waylon Li","Yftah Ziser"],"outgoing":[{"source":"arxiv:2602.06130","target":"arxiv:2403.07714","evidence":"id","id":"arxiv:2403.07714","title":"StableToolBench: Towards Stable Large-Scale Benchmarking on Tool Learning of Large Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ed03ae4fed851402f7c4c5378410fc34374148218bc20e4e7e92522063d06df9/"},{"source":"arxiv:2602.06130","target":"arxiv:2406.03689","evidence":"title","id":"arxiv:2406.03689","title":"Evaluating the World Model Implicit in a Generative Model","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/"}],"incoming":[],"text_parts":[{"start":0,"end":15528,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html"},{"start":15528,"end":31145,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html"},{"start":31145,"end":46866,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html"},{"start":46866,"end":60826,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html"}],"contents":[{"title":"Self-Improving World Modelling with Latent Actions","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Abstract","offset":54,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Keywords:","offset":1622,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"1 Introduction","offset":1640,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"2 Related Work","offset":6195,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Intrinsic World Models.","offset":6214,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Self-Improving Learning.","offset":7587,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"3 Methodology","offset":9449,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"3.1 Task Formulation","offset":9467,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"3.2 SWIRL","offset":10887,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Intuition.","offset":10902,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Objectives.","offset":11868,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"3.3 Theoretical Analysis","offset":14211,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Phase I: FWM optimisation via Variational Information Maximisation.","offset":14554,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Theorem 3.1 (FWM Lower Bound) .","offset":15168,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Proof.","offset":15513,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-1.html#text"},{"title":"Phase II: IDM optimisation via ELBO Maximisation.","offset":18235,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Theorem 3.2 (IDM Lower Bound) .","offset":18703,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Proof.","offset":19026,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"4 Experiments and Results","offset":20796,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"4.1 Experimental Setup","offset":20826,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Models.","offset":20854,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"SFT Warm-up.","offset":21315,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Iterative RL.","offset":22259,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Benchmarks.","offset":23216,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Baselines.","offset":24489,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Evaluation Metrics.","offset":25544,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"4.2 Single-turn Visual Dynamics Prediction","offset":25888,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Aurora-Bench .","offset":25936,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"ByteMorph .","offset":27450,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"4.3 Multi-turn Visual Dynamics Prediction","offset":28182,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"WorldPredictionBench .","offset":28229,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"4.4 Textual Environments","offset":29415,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"4.5 Analysis","offset":30337,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Iterative RL.","offset":30355,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"RL vs SFT.","offset":31128,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-2.html#text"},{"title":"Sharing θ \\theta and ϕ \\phi .","offset":32274,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"5 Conclusion","offset":33419,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Impact Statement","offset":34305,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Acknowledgements","offset":35885,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Appendix A Theoretical Derivation","offset":36400,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"A.1 Phase I: Learnability of the Forward Model (FWM)","offset":37602,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Theorem A.1 .","offset":38469,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Proof.","offset":38759,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"A.2 Phase II: Learnability of the Inverse Model (IDM)","offset":41284,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Theorem A.2 .","offset":41751,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Proof.","offset":42039,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Appendix B Implementation Details","offset":43956,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"B.1 Hyperparameters","offset":43994,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"B.2 Evaluation Prompt for GPT-4o as a Judge","offset":46311,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Appendix C Sanity Check Results for General Image Editing.","offset":46803,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-3.html#text"},{"title":"Appendix D IDM Analysis","offset":49180,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Aurora-Bench Evaluation.","offset":49208,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Interpretability of Latent Actions and Reward Hacking.","offset":50058,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Appendix E Inference-time Verification","offset":51688,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Appendix F Detailed Results for Comparing SFT and SWIRL .","offset":52664,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Appendix G Detailed Results for WorldPredictionBench","offset":54034,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Appendix H Detailed Results for Iterative Results","offset":55226,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Appendix I Ablation on GRPO Rollout Size.","offset":56835,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Appendix J Qualitative Examples","offset":57867,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"},{"title":"Instructions for reporting errors","offset":59862,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/part-4.html#text"}],"text_code_points":60826,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/source.md"}