Adam Safron, Michael Levin, Victoria Klimaj, Zahra Sheikhbahaee, Dalton Sakthivadivel, Adeel Razi, et al. (2026), Philosophical Transactions of the Royal Society A 384(2320), 20240533 (introduction to the theme issue “World models in natural and artificial intelligence”) — DOI-10.1098/rsta.2024.0533
Key takeaways
- “World model” is used in senses that range from any information a system cognizes over, to explicit causal structure, to causal prediction in the service of agency. LLMs clearly meet only the first, and the editors argue that world models are context-dependent constructions shaped by embodiment, goals and social learning, not neutral abstractions.
Seeded from the abstract and a skim, before the reading in NOTE-136. What follows is what the work says about itself.
This editorial introduces a theme issue on how natural and artificial intelligences model the world, and what that reveals about cognition and the continuity of life and mind. The editors take no single definition. They survey causal, self-referential, goal-directed, collective and narrative forms of world modelling. A recurring question is whether AI trained on vast data captures the context-sensitive, temporally embedded, value-laden aspects of biological world modelling, or mainly exploits statistical surface regularities. Contributions also address embodiment, complexity, learning architectures and the social settings of world models. The aim is to clarify the conceptual landscape and to frame questions about grounded understanding, robust agency and human-like general intelligence.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 15 seconds of active reading over 1 session. It was filed Deferred, unread. NOTE-136 is the close reading of 2026-09-26, and it placed the work: Active — worth reading as the map of the theme issue and a short, citable statement of the three senses of “world model”. It argues nothing itself. Its claims are editorial glosses on 17 contributions, which carry the evidence.
When seeded, before the reading in NOTE-136, the entry gave a priority for a deeper reading: low — t=15 s. As an editorial it is fully captured by the skim, and its value is as a map to the theme issue’s papers.
What the seeding asked a deeper reading to check (NOTE-136 is that reading):
- The three-tier criterion for “world model” in §3 is a usable distinction for any note on whether LLMs have world models, and it could anchor a nucleation cluster on grounding and understanding (compare LIT-131 Freeborn and LIT-207 Roberts).
- It is an editorial: it surveys and frames rather than argues. The individual contributions (Hofstadter; Krakauer et al.; Ku et al.) are where the claims live and may deserve their own notes.
- The editorial is by the issue’s guest editors and co-organisers; the author list (14) is broad and several listed authors contributed “conceptualization” only, per the contributions statement.
Access when seeded: royalsocietypublishing.org returned 403 to every request (article page, DOI and PDF). Read the full open-access HTML of the editorial from the Wayback Machine capture of 2026-05-19: abstract, keywords, §1 “World models in artificial and natural intelligence”, the subsection titles of §2 “Summary of contributions and commentary” and its opening Hofstadter summary, and §3 “Conclusion and future directions”. Received 9 Jan 2026, accepted 12 Jan 2026, published online 14 May 2026 per the page and Crossref. Adam Safron’s site confirms it is the issue’s introductory editorial.