{"id":"arxiv:2406.03689","title":"Evaluating the World Model Implicit in a Generative Model","date":"2024-06-06","section":"Reasoning and the \"physics\" of language models","quality":0.186581,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/","source_url":"https://arxiv.org/abs/2406.03689","text_available":true,"authors":["Keyon Vafa","Justin Y. Chen","Ashesh Rambachan","Jon Kleinberg"],"outgoing":[],"incoming":[{"source":"arxiv:2602.06130","target":"arxiv:2406.03689","evidence":"title","id":"arxiv:2602.06130","title":"Self-Improving World Modelling with Latent Actions","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac56340e6b81e232b48b92dcf79fe37efc1c00e676413e55c9d119a809eb644b/"}],"text_parts":[{"start":0,"end":15462,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html"},{"start":15462,"end":30657,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html"},{"start":30657,"end":46051,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html"},{"start":46051,"end":61663,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html"},{"start":61663,"end":66143,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-5.html"}],"contents":[{"title":"Evaluating the World Model Implicit in a Generative Model","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"Abstract","offset":61,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"1 Introduction","offset":1193,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"2 Framework","offset":9017,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"2.1 Recovering world models","offset":11537,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"Definition 2.1 .","offset":11899,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"Definition 2.2 .","offset":12495,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"Proposition 2.3 .","offset":12828,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"2.2 Next-token prediction is a fragile metric for recovering structure","offset":13395,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"2.3 The Myhill-Nerode interior and boundary","offset":15413,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-1.html#text"},{"title":"Definition 2.4 .","offset":16278,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"2.4 Compression and distinction metrics for evaluating world models","offset":17720,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"Definition 2.5 .","offset":17934,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"Definition 2.6 .","offset":19163,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"3 Illustration: Do Transformers Recover the Street Map of New York City?","offset":21936,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"3.1 Data and models","offset":22918,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"3.2 Evaluating world models","offset":26243,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"3.3 Reconstructing implicit maps","offset":30256,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-2.html#text"},{"title":"3.4 Implication of failing to recover the world model: detour fragility","offset":33185,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"4 Other Applications: Othello and Logic Puzzles","offset":34570,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"5 Conclusion","offset":38659,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"Acknowledgements","offset":39881,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"Appendix A Proof","offset":40482,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"Proof of Proposition 2.3 .","offset":40503,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"Appendix B Reconstructed Maps","offset":42573,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"B.1 Algorithm","offset":42834,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"B.2 Maps","offset":44955,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-3.html#text"},{"title":"Appendix C Deterministic Finite Automata and Myhill-Nerode","offset":48067,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"Definition C.1 (Equivalent sequences) .","offset":48683,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"Theorem C.2 ( [ 26 , 27 ] ) .","offset":49020,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"Appendix D Additional results","offset":49494,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"Appendix E Evaluation metric details and ablations","offset":52050,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"E.1 Implementation details","offset":52376,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"E.2 Ablations","offset":59153,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"Appendix F Rides data construction and training","offset":61611,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-4.html#text"},{"title":"Appendix G Additional maps","offset":65148,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-5.html#text"},{"title":"Instructions for reporting errors","offset":65179,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/part-5.html#text"}],"text_code_points":66143,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/41da4da8a33d9ab26af9650748067540eaf4cd386071ae0224a06450cdbb1fd0/source.md"}