nucleation

LIT-1891: General agents contain world models

1 min read

Properties5
StatusActive — read in full 2026-10-31 (NOTE-1469); worth reading for the cleanest theorem that competence across a family of goals fixes the environment's dynamics, and for exactly what "contains a world model" means in it: the transition function is computable from the goal-conditioned input-output map (THEORY-1249), which holds of a lookup table as much as of a planner (THEORY-1260). It is the objection to the functional reading of CLAIM-519, and it does not meet that claim's defeated_if.
Version2
Tagsagency, learning-theory, anthology-candidate
Cited as “paper” byNOTE-1469 — Richens, Abel, Bellot & Everitt — general agents contain world models
Cited as “source” byTHEORY-1249 — In a fully observed, finite, communicating, stationary controlled Markov process, a deterministic goal-conditioned policy that achieves every composite goal of depth up to n with probability at least 1 − δ times the optimum determines every transition probability to within √(2P(1−P)/((n−1)(1−δ))), read from its first action on either-or goals; for a policy competent only at one-step goals no bound follows, THEORY-1260 — Richens et al.'s sense of "an agent contains a world model" is that the transition function can be computed from the agent's goal-conditioned input-output map, uniquely to within the theorem's error, so the attribution is not a stance an observer may withhold, unlike a representation-theorem fit; but it holds of every implementation of the same map, a lookup table included, and so cannot show that a predictive component is realized inside the agent

Graph View

Richens, J., Abel, D., Bellot, A., & Everitt, T. (2025). General agents contain world models. Proceedings of the 42nd International Conference on Machine Learning, PMLR 267. arXiv:2506.01622.

Read in full in NOTE-1469.

Why this is here

Not named in the exchange. It is the published argument against CLAIM-519, the owner’s U20 thesis that a world model is an explanatory attribution. Read with Block’s lookup-table argument (THEORY-092). The reading settles how far it reaches. Under its assumptions the transition function is identifiable from the goal-conditioned policy, uniquely to within the bound (THEORY-1249). So on the functional reading, ascribing a world model to a goal-general agent is not a stance one may withhold, and the thesis’s “folk explanatory tool” is too weak there. But the recovery map runs from the policy, not from internals, and holds of a lookup table over histories and goals (THEORY-1260). It makes no resource comparison and no intervention, so it does not meet the thesis’s defeated_if and does not reach its claim about realized mechanism. On the thesis’s own setup, one fixed interface with no goal input, Theorem 2 shows nothing need be encoded at all.

Backlinks