World model
A world model is a model of how an environment changes in response to what happens in it, kept by a learning system so that it can try actions in imagination before trying them in the world. The idea comes from the psychology of the 1940s, and since 2018 the phrase has covered a wide range of machine-learning systems, from the networks that let a game-playing program plan to video generators that their makers call simulators of the physical world. Whether the larger systems hold a model of the world or only statistics that resemble one is still open, and it has been tested on, among other things, the board game Othello.
Origins
The idea is older than computing. The Scottish psychologist Kenneth Craik wrote in The Nature of Explanation (1943) that an organism which carries "a small-scale model" of external reality and of its own possible actions can try out alternatives before it acts.[1] In 1991 Richard Sutton described Dyna, an architecture in which an agent learns a model of its environment from experience and uses the model to plan, so that planning and reacting share one learner.[2]
The phrase took its present sense from "World Models", a paper of 2018 by David Ha and Jürgen Schmidhuber. They trained generative networks on reinforcement-learning environments, then used features from the model as input to a small policy, and trained an agent entirely inside "its own hallucinated dream" before transferring the policy back to the real environment.[3]
Designs
| System | Year | What the model predicts | Result reported |
|---|---|---|---|
| Dyna[2] | 1991 | The next state and reward | Planning steps on the learned model speed up learning |
| World Models[3] | 2018 | A compressed code of the next frame | Policies trained inside the model transfer to the real task |
| MuZero[4] | 2020 | Policy, value and reward, with no model of the screen | Matched AlphaZero in Go, chess and shogi without being given the rules |
| DreamerV3[5] | 2023 (preprint) | Latent states and rewards, used to imagine futures | One configuration over more than 150 tasks; reported as the first to collect diamonds in Minecraft without human data |
| Genie[6] | 2024 | The next video frame, given a latent action | 11 billion parameters; trained on unlabeled internet video |
| V-JEPA 2[7] | 2025 | The representation of hidden video | Pretrained on over a million hours of video, then used for robot planning |
MuZero's model never reconstructs the screen. It learns the quantities that the search requires, which are the action probabilities, the value of the position and the reward.[4] The Dreamer family imagines trajectories in a learned latent space, an approach that builds on a 2018 paper on planning from pixels.[8]
Prediction without reconstruction
In a position paper of 2022, Yann LeCun proposed an architecture for autonomous machines in which a world model predicts in representation space and not in pixels, since most of the detail in a frame cannot be predicted and a model that tries spends its capacity on it.[9] I-JEPA, one of the designs that followed, predicts the representations of hidden blocks of an image from one visible block, with no generative decoder.[10] The same idea has been applied to electrocardiograms,[11] and see joint embedding predictive architecture for the design in detail.
Do they model the world?
Video generators have been called simulators. OpenAI's technical report on Sora, in February 2024, concluded that "scaling video generation models is a promising path towards building general purpose simulators of the physical world".[12] A direct test needs a world whose rules are fully known, such as a board game.
One test used Othello. A language model of the GPT type was trained only to predict legal moves, without being told the rules or that there is a board, and it developed an internal representation of the board state that experimenters could read out and change; changing it altered the moves the network predicted.[13] A second study treated worlds with fixed rules as finite automata and scored how well generative models recovered them, with metrics drawn from the Myhill-Nerode theorem. In navigation, game-playing and logic puzzles the models did well on the usual diagnostics. The new metrics showed their world models to be "far less coherent than they appear", and a model asked for a slightly different task could fail.[14]
At Molecular Reality
Molecular Reality Corporation proposes a world model for nanopore signals. On its RLMF page the company describes a "public domain nanopore world model" that "will learn to understand signal-producing events in nanopore current data at the level of the sensing zone and instrument parameters", and would grow into "a comprehensive molecular sensory organ for all molecular sciences".[15] The program is called RLMF, and the intended central model is Maxine's Mind.[16] The company's field guide files the idea among its plans: "Both are proposals, not products, and I want that said first."[17]
Its predictor is a JEPA, a design proposed by researchers at Meta. The critique's reason for predicting in a latent space is that a model reconstructing every sample of a 200 kHz trace would spend "~95% of its capacity memorizing noise".[18] In June 2026 the company published a critique of its own plan, written under the name of its mascot Keltar, that endorsed that one decision and questioned most of the others. The critique is described in The Nanopore World Model.
See also
References
- ^ Craik, K. J. W. (1943). The Nature of Explanation. Cambridge University Press. p. 61.
- ^ a b Sutton, R. S. (1991). "Dyna, an integrated architecture for learning, planning, and reacting". ACM SIGART Bulletin 2 (4): 160-163. doi:10.1145/122344.122377.
- ^ a b Ha, D.; Schmidhuber, J. (2018). "World Models". arXiv. doi:10.48550/arXiv.1803.10122.
- ^ a b Schrittwieser, J.; et al. (2020). "Mastering Atari, Go, chess and shogi by planning with a learned model". Nature 588 (7839): 604-609. doi:10.1038/s41586-020-03051-4.
- ^ Hafner, D.; et al. (2025). "Mastering diverse control tasks through world models". Nature 640 (8059): 647-653. doi:10.1038/s41586-025-08744-2.
- ^ Bruce, J.; et al. (2024). "Genie: generative interactive environments". arXiv. doi:10.48550/arXiv.2402.15391.
- ^ Assran, M.; et al. (2025). "V-JEPA 2: self-supervised video models enable understanding, prediction and planning". arXiv. doi:10.48550/arXiv.2506.09985.
- ^ Hafner, D.; et al. (2018). "Learning latent dynamics for planning from pixels". arXiv. doi:10.48550/arXiv.1811.04551.
- ^ LeCun, Yann (June 27, 2022). "A path towards autonomous machine intelligence". OpenReview.
- ^ Assran, M.; et al. (2023). "Self-supervised learning from images with a joint-embedding predictive architecture". arXiv. doi:10.48550/arXiv.2301.08243.
- ^ Weimann, K.; Conrad, T. O. F. (2024). "Self-supervised pre-training with joint-embedding predictive architecture boosts ECG classification performance". arXiv. doi:10.48550/arXiv.2410.13867.
- ^ "Video generation models as world simulators". OpenAI. February 15, 2024.
- ^ Li, K.; et al. (2022). "Emergent world representations: exploring a sequence model trained on a synthetic task". arXiv. doi:10.48550/arXiv.2210.13382.
- ^ Vafa, K.; et al. (2024). "Evaluating the world model implicit in a generative model". arXiv. doi:10.48550/arXiv.2406.03689.
- ^ "Reinforcement Learning with Molecular Feedback". epicquest.bio. Molecular Reality Corporation. 2026.
- ^ "Company records". Molecular Reality Corporation. October 2026.
- ^ "Keltar's Field Guide to Molecular Reality". Molecular Reality Corporation. August 2026.
- ^ "Keltar's Critique v2: The Nanopore World Model (Hardened)". Molecular Reality Corporation. June 2026.