World Models

A world model is a learned or constructed internal model that predicts how the state of an environment changes over time and in response to an agent's actions, and that can be used to plan, imagine outcomes or interpret observations. In AI research the term covers explicit latent-dynamics models used in reinforcement learning, video- and interaction-generating models, and the question of whether large language models acquire implicit models of the processes that generated their training data.

Planning requires predicting consequences. An agent that can predict what will happen if it acts can evaluate options without executing them, which reduces the number of costly or dangerous real-world trials. Proposals for general machine intelligence frequently place a world model at the centre of the architecture (LeCun, 2022), and AGI-PATH-001 lists world-model approaches as one of several pathways. Whether a world model is necessary for general intelligence is Contested (see AGI-REQ-001); this article describes what is meant by the term and what has been demonstrated.

Three related meanings are often conflated.

1. Control-oriented world models. A model of dynamics learned so that an agent can plan or train a policy in imagination (Ha & Schmidhuber, 2018; Hafner et al., 2025).
2. Generative environment models. Models that produce video or interactive environments conditioned on text, images or actions (Bruce et al., 2024).
3. Implicit world models in sequence predictors. The hypothesis that a model trained only to predict the next token or move develops internal structure that mirrors the underlying world (Li et al., 2023; Gurnee & Tegmark, 2023; Vafa et al., 2024).

The idea of an agent with an internal model of its environment is long-standing in cognitive science and control theory. In the neural-network literature, Ha and Schmidhuber (2018) presented a system in which a generative model of the environment is learned from observations and a small controller is trained within the model's own simulated dynamics. The same period saw model-based reinforcement learning advance in board games: MuZero learned a model sufficient for planning without being given the rules of the game, and reached strong performance on Go, chess, shogi and Atari (Schrittwieser et al., 2020).

Hafner et al. (2025) introduced DreamerV3, which uses a learned world model with a single fixed configuration across many tasks, and which the authors state is the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula. LeCun (2022) proposed an architecture for autonomous agents built around a world model trained with joint-embedding predictive methods rather than pixel-level generation. Generative video and interactive-environment models followed: Genie is described by its authors (at the date of the 2024 paper) as an 11-billion-parameter model that "can be considered a foundation world model" trained on unlabelled Internet video that can generate action-controllable virtual worlds from prompts (Bruce et al., 2024). V-JEPA 2 extended joint-embedding methods to video and, with a small amount of robot data, to planning in robot manipulation (Assran et al., 2025).

In parallel, interpretability studies asked whether sequence models already contain world models. Li et al. (2023) trained a GPT-style model on Othello move sequences and found internal board-state representations that could be intervened on. Gurnee and Tegmark (2023) reported linear representations of space and time in Llama-2 models. Vafa et al. (2024) proposed metrics based on the Myhill-Nerode theorem and found that models which do well on standard tests had much less coherent implicit world models than those tests suggested.

#4.1 Formal structure

A world model is commonly formalised as a stochastic transition model over latent states. Given observations $o_t$, actions $a_t$ and latent state $z_t$, a typical model comprises:

• an encoder, $z_t \sim q_\phi(z_t \mid z_{t-1}, a_{t-1}, o_t)$;
• a dynamics (prior) model, $\hat{z}_t \sim p_\theta(\hat{z}_t \mid z_{t-1}, a_{t-1})$;
• a decoder or predictive head, $\hat{o}_t \sim p_\theta(\hat{o}_t \mid z_t)$, and a reward predictor $\hat{r}_t$.

Training maximises a variational objective that balances reconstruction or prediction accuracy against closeness between the posterior and prior over latent states. In this family the learned dynamics can be rolled forward from a starting state to produce imagined trajectories, and a policy is optimised on those trajectories (Ha & Schmidhuber, 2018; Hafner et al., 2025). This describes the general design; the specific objectives, architectures and normalisation tricks differ by system.

#4.2 Planning with a learned model

Given a model, an agent chooses actions by search or optimisation over imagined futures. MuZero combines a learned model of reward, value and policy-relevant quantities with Monte Carlo tree search (Schrittwieser et al., 2020): the learned model need not reconstruct observations, only to predict what matters for planning. This shows that a world model, in the sense of a model useful for decisions, need not be a faithful simulator.

#4.3 Prediction in representation space

LeCun (2022) argues for predicting in an abstract representation space instead of predicting every detail of the input, because many details of the world are not predictable. V-JEPA 2 follows this approach for video: the authors report pre-training on over 1 million hours of internet video and images, and a robot-planning variant post-trained on fewer than 62 hours of unlabelled robot video, deployed zero-shot on robot arms in two laboratories for pick-and-place tasks (Assran et al., 2025). These are results reported by the developing group and are Preliminary.

#4.4 Implicit world models

For language and sequence models, the claim is that internal activations encode latent variables of the data-generating process. Evidence takes the form of probes (decoding the variable from activations) and interventions (changing the representation and observing changes in output). Li et al. (2023) found a nonlinear board representation in the Othello setting; Gurnee and Tegmark (2023) found linear encodings of geographic and temporal variables in Llama-2 models. Vafa et al. (2024) tested coherence on game playing, logic puzzles and navigation, and found that the implicit models were much less consistent than standard tests indicated, which can cause failures on closely related tasks. Probing results show that information is present and decodable; they do not by themselves show that the model uses a coherent model for inference.

A further classification concerns explicit versus implicit models, deterministic versus stochastic dynamics, and observation-space versus latent-space prediction. The choice among these determines the computational cost of rollouts and the kinds of errors the model compounds.

• Game playing and simulated control. MuZero and DreamerV3 are demonstrations in simulated or game environments, not deployed products (Schrittwieser et al., 2020; Hafner et al., 2025).
• Generative environments. Genie generates playable worlds from prompts; the authors also describe using its learned latent action space to train agents to imitate behaviours from unseen videos (Bruce et al., 2024). This is a laboratory demonstration.
• Robot planning. V-JEPA 2-AC, per its developers, performed zero-shot pick-and-place on Franka arms in two laboratories without data collected from those robots (Assran et al., 2025). This is a laboratory result reported by the developers; it has not been shown here to extend to open-ended tasks. For the broader field see TECH-ROBOT-001.
• Interpretability studies. The Othello and Llama-2 results (Li et al., 2023; Gurnee & Tegmark, 2023) are used as evidence in debates about what language models represent. See FOUND-014 for the related philosophical question.

Advantages. Planning in a model can make learning more sample-efficient than purely model-free learning; a model supports counterfactual evaluation of actions; and a single model can serve multiple tasks. (The claim of sample efficiency is an expected benefit of the approach; the sources reviewed here demonstrate performance but do not compare data efficiency in every case.)

Limitations.

• Compounding error. Small prediction errors accumulate over long rollouts, limiting the planning horizon.
• Stochasticity and unpredictability. Many aspects of the world cannot be predicted in detail, which motivates prediction in latent space (LeCun, 2022) but leaves open how to choose what to predict.
• Evaluation difficulty. A model may look plausible while being physically or logically inconsistent. Vafa et al. (2024) show standard task accuracy can overstate the coherence of implicit models.
• Generative realism is not understanding. A model that produces visually convincing video need not have an accurate model of physics. The sources reviewed here describe generation, not physics-accurate prediction.
• Scale and cost. Training large video models is computationally expensive; see ECON-COMP-001.

Open questions.

1. Is a learned world model necessary for general intelligence, or can language-based reasoning over text and tools substitute for it? Contested.
2. Do language models contain usable world models, or only decodable correlates? Contested.
3. How should world-model quality be measured so that tests catch inconsistency rather than reward surface fluency?
4. Can world models learned from passive video support planning in the real world at scale?

As of 2026-10-19 the sources reviewed support the following. Learned world models enable strong performance in simulated and game environments (Schrittwieser et al., 2020; Hafner et al., 2025), on the authors' reports. Generative interactive models and video-based predictive models exist as research systems (Bruce et al., 2024; Assran et al., 2025). Probing studies find structured internal representations in language and sequence models (Li et al., 2023; Gurnee & Tegmark, 2023), while a metric-based study finds substantial incoherence in implicit models (Vafa et al., 2024). The overall picture of what implicit world models LLMs have is Contested. Figures such as parameter counts, training-data sizes and benchmark scores in the papers above are given at the dates of those papers; for current systems consult the Living Reference collections (MODEL-001, FACT-001). The architectures and training data of closed commercial video and world-model systems are, in general, undisclosed, and this article makes no claims about them.

This section is descriptive and registers no forecast.

• Near term (0 to 3 years). Expect continued work on video-based predictive models, interactive-environment generators and benchmarks for physical and causal consistency.
• Medium term (3 to 10 years). Whether world-model approaches provide a path to general capability, or are one component alongside language-based reasoning and tools, will depend on evidence that does not yet exist in the sources reviewed.
• Long term (10+ years). Speculative; connections to embodied systems and scientific simulation are discussed in TECH-ROBOT-001 and APP-SCI-001.

• Foundational Concepts:
  • AGI-PATH-001 Pathways to AGI: World-model approaches among proposed routes to AGI.
  • TECH-RL-001 Reinforcement Learning: Model-based reinforcement learning.
  • AGI-REQ-001 Requirements for General Intelligence: World models as a candidate requirement.
• Related Methods & Technologies:
  • TECH-DIFF-001 Diffusion Models: A common basis of generative video models.
  • TECH-ROBOT-001 Robotics and Embodied AI: Use of world models for robot planning.
  • AGI-CL-001 Continual Learning and Persistent Memory: Keeping a world model up to date.
• Advanced Implications & Future Trajectories:
  • FOUND-014 Consciousness and Machine Understanding: Whether internal models amount to understanding.
  • AGI-TEST-001 Testing Claims of General Intelligence: Testing coherence of world models.