Concepts

What are world models? The bet on AI that learns how the world works.

July 14, 2026 / Rudy van den Brink / ~18 min read

If you follow AI research at all, you have heard the term world model a lot lately. Google DeepMind's Genie 3 generates interactive environments you can walk around in. Meta's V-JEPA 2 watched a million hours of video and then controlled a robot arm it had never seen. NVIDIA built an entire platform (Cosmos) around the idea, Fei-Fei Li's World Labs raised hundreds of millions to pursue "spatial intelligence", and Yann LeCun left Meta after twelve years to found a startup dedicated to the concept — reportedly raising over a billion dollars before shipping anything.

The pitch behind all of this goes roughly as follows. Large language models learn the statistics of text, and text is a thin, lossy description of reality. A world model instead learns the dynamics of an environment — how the state of the world evolves, and how actions change it — from video and interaction. Because such a model captures something closer to the causal structure of the physical world, the argument goes, it can support prediction and planning over much longer horizons than a system that generates one token at a time and hopes for the best.

A claim this popular deserves a careful look. In this article I want to do three things: pin down the terminology (one term in particular gets misused constantly, and I'll explain why it matters), lay out what world models actually are and where the long-horizon argument comes from, and then — most importantly — look at what the empirical evidence says. The short version: the diagnosis of LLM limitations is largely sound, the there is theoretical motivation for world models, and the evidence that current systems actually learn physics is considerably weaker than the marketing suggests.

First, the terminology: "model collapse" is something else

A common framing of the motivation — one I've heard in many conversations — is that LLMs suffer "model collapse" when reasoning across many steps. The failure is real, but that isn't what model collapse means, and the distinction is worth getting right because the two problems have entirely different causes and fixes.

Model collapse is a training-time phenomenon: when generative models are trained on data produced by earlier generative models, rare events disappear from the learned distribution and output diversity degrades over successive generations (Shumailov et al., 2024). It is about data provenance, not reasoning. The failure of LLMs on long reasoning chains is an inference-time phenomenon, described in the literature as compounding errors (Dziri et al., 2023) or accuracy collapse (Shojaee et al., 2025).

With that sorted, the inference-time problem itself is well documented, and it is the one that matters for the world-model story.

How LLMs actually fail on long reasoning chains

The cleanest study of the mechanism is Faith and Fate: Limits of Transformers on Compositionality (Dziri et al., 2023). The authors formalize multi-step problems — multi-digit multiplication, logic grid puzzles, a classic dynamic programming task — as computation graphs, and measure how transformer performance varies with the size and depth of the graph. The pattern is consistent: accuracy is high when the required computation is shallow, then decays rapidly toward zero as the number of dependent steps grows. Two findings stand out. First, errors made early in a chain propagate: a small per-step error rate compounds across dependent steps. Second, much of the apparent reasoning success is what the authors call linearized subgraph matching — the model succeeds when large portions of the required computation resemble patterns it saw in training, which is pattern reuse rather than general procedure execution.

Apple's 2025 paper The Illusion of Thinking (Shojaee et al., 2025) extended this to dedicated reasoning models — the ones that generate long chains of thought before answering. Using puzzles with controllable complexity (Tower of Hanoi, river crossing, blocks world), they found that accuracy doesn't degrade gracefully as problems get harder: it holds up to a complexity threshold and then collapses to zero, and near the threshold the models paradoxically reduce their reasoning effort, as if giving up. I should note this paper is contested — a widely circulated rebuttal (Opus & Lawsen, 2025) argued that part of the collapse is an artifact of output-token limits and task encoding — but the qualitative finding is consistent with independent work. On PlanBench, a benchmark of simple symbolic planning problems, OpenAI's o1 scored 97.8% on standard Blocksworld puzzles but fell to 37.3% on a semantically obfuscated version of the same underlying problems, and pre-reasoning LLMs scored below 5% (Valmeekam et al., 2024). If the models were executing a general planning procedure, renaming the objects wouldn't matter.

Yann LeCun has long argued that this is baked into the autoregressive setup: if each generated token carries some independent probability of error, the probability that a long chain stays on the rails decays exponentially — at a 1% per-token error rate, a 100-step chain is correct only about 37% of the time (0.99100 ≈ 0.366). The clean mathematical version of this argument is disputed — errors in real models are not independent, and models can and do self-correct mid-chain (see Arbuzov et al., 2025, for a careful treatment) — but as an empirical summary of the benchmark literature, "reliability degrades with the length and novelty of the required reasoning chain" is well supported.

So the diagnosis half of the premise holds up, with one refinement: the failure isn't purely about horizon length. It's about horizon length combined with novelty — chains that can't be stitched together from familiar patterns. That refinement matters later, because it raises the bar for what a replacement architecture must demonstrate.

What a world model is

A world model is an internal, learned model of how an environment evolves: given the current state and (crucially) an action, predict the next state. Once you have such a model, you can do something qualitatively different from reacting: you can try actions in imagination, see where they lead, and pick the sequence that leads somewhere good — before touching the real world.

The idea is much older than the current wave. The psychologist Kenneth Craik proposed in 1943 that organisms carry a "small-scale model" of external reality in their heads, used to try out alternatives before acting — and modern neuroscience has found substantial evidence for internal forward models in, for example, motor control, where the brain predicts the sensory consequences of its own commands (Wolpert et al., 1995). In machine learning, Richard Sutton's Dyna architecture (1991) made the same move formal: learn a model of the environment from experience, then use the model to generate simulated experience for planning. This became the field of model-based reinforcement learning, distinguished from model-free RL precisely by whether the agent learns environment dynamics or only a policy.

The modern deep-learning formulation arrived with Ha and Schmidhuber's 2018 paper, literally titled World Models. They trained a variational autoencoder to compress game frames into a small latent vector, a recurrent network to predict the next latent state, and then — the memorable part — trained a controller entirely inside the learned model's own "dream", transferring the resulting policy back to the real game successfully. Nearly everything in today's world-model literature is a descendant of that loop. Three lineages matter most:

  • Latent world models for control. DeepMind's MuZero (Schrittwieser et al., 2020) learned a model that predicts only what planning needs — reward, value, and policy — and reached superhuman play in Go, chess, shogi, and Atari without being told the rules. The Dreamer line learns behaviors inside a learned latent model; DreamerV3 (Hafner et al., 2025) mastered over 150 tasks with one fixed configuration and was the first agent to collect diamonds in Minecraft from scratch, with no human data or curriculum — a genuinely long-horizon task.
  • Generative video as world simulation. OpenAI's technical report for Sora was titled "Video generation models as world simulators", and DeepMind's Genie line makes the simulation interactive: Genie 3 (DeepMind, 2025) generates explorable environments at 720p and 24 frames per second that stay consistent for minutes. Wayve's GAIA-1 (Hu et al., 2023) does the same for driving scenes; the Cosmos platform (NVIDIA, 2025) packages video world models as infrastructure for robotics and autonomous-vehicle training.
  • Prediction in representation space. LeCun's 2022 position paper, A Path Towards Autonomous Machine Intelligence, argues that predicting raw pixels wastes capacity on irrelevant, unpredictable detail (every leaf on a tree), and that the model should instead predict in an abstract embedding space that keeps what is predictable and discards the rest. That is the JEPA family (Joint Embedding Predictive Architecture). V-JEPA 2 (Assran et al., 2025) is the current flagship: pretrained on over a million hours of internet video, then given about 62 hours of robot interaction data to become action-conditioned — after which it could plan pick-and-place manipulation on a real robot arm in labs it had never seen, zero-shot.

The term "world model" is doing a lot of work across these three lineages, and it pays to notice which sense an author means: a dynamics model used for planning (MuZero, Dreamer, JEPA), a generative video simulator (Sora, Genie, Cosmos), or — a fourth sense — the implicit model of the world a plain LLM might acquire as a byproduct of next-token prediction. Claims true of one sense are routinely, and wrongly, transferred to the others. World Labs' Marble is different again: it generates persistent, navigable 3D scenes — spatial structure rather than temporal prediction.

Where the long-horizon argument comes from

Now to the second half of the premise: why would a world model predict further ahead than an LLM can reason? The honest answer is that the advantage, where it exists, does not come from world models being immune to error accumulation. They aren't — roll any learned dynamics model forward autoregressively and small prediction errors compound into drift, exactly as they do in token space. Genie 3's headline achievement was staying coherent for minutes; that this counts as a breakthrough tells you how hard the horizon problem remains inside the world-model paradigm too. The argument rests instead on three more specific mechanisms.

Closed-loop planning instead of open-loop generation. An LLM writing a chain of thought commits to every step it generates: each token becomes context for the next, so an early mistake contaminates everything downstream. A planner using a world model works differently — imagine several candidate action sequences, score the predicted outcomes, execute only the first action, observe what actually happened, and replan from the true new state. Imagined futures are disposable; reality corrects the model at every step. This is model-predictive control, and it converts one long open-loop prediction problem into many short ones. Long-horizon competence then doesn't require long-horizon prediction accuracy — which is how DreamerV3 can pull off a task as deep as mining a Minecraft diamond while its world model imagines only short rollouts.

Abstraction. Predicting the right things is easier than predicting everything. In pixel space or token space, most of the future is unpredictable detail; in a well-chosen abstract state space, dynamics can be nearly deterministic ("the cup will fall") even when appearances aren't ("exactly how it shatters"). This is the core JEPA claim: per-step errors shrink when the model predicts abstract state, so usable horizons stretch.

Grounding in action. A model trained on observation and action can learn counterfactuals — what would happen if — which is the raw material of causal knowledge. Text describes the world's regularities secondhand; interaction data contains them directly.

There is also a genuinely elegant theoretical result here. Richens and Everitt (2024) proved that any agent that adapts successfully to a sufficiently large set of distributional shifts must have learned an approximate causal model of its environment — there is no model-free shortcut to robustness. A follow-up, General agents need world models (Richens, Everitt & Abel, 2025), showed the same for goal-directed behavior: any agent meeting a performance bound across sufficiently many long-horizon goals must have learned a predictive model of the environment, and the required accuracy of that model increases with the task horizon. You can even extract the world model from the agent's policy. Two implications follow. The optimistic one: the world-model research program is aimed at something provably necessary, not merely fashionable. The subtler one: "LLMs versus world models" is a false dichotomy — to the extent an LLM-based agent really does generalize across long-horizon tasks, it must contain an implicit world model already. The question is never whether to have one, but whether explicit training objectives produce a better one than next-token prediction does as a side effect.

LLM, chain-of-thought World model + planner
What is learned Statistics of token sequences Environment dynamics: state × action → next state
Prediction space Discrete tokens (input space) Latent state (Dreamer, JEPA) or pixels (video models)
Action-conditioned Not natively Yes — that's the point
Inference regime Open loop: each step becomes context for the next Closed loop: imagine, act, observe, replan
Long-horizon failure mode Compounding errors in the committed chain Rollout drift — mitigated by replanning from observed state

Do they actually learn physics? The evidence

This is where the story gets uncomfortable for the strong version of the pitch. The claim that watching enough video teaches a model "generative properties of the physical world" is an empirical claim, and it has now been tested several times with careful methodology. The results are consistent, and they are not flattering.

Visual realism is not physical understanding. Physics-IQ (Motamed et al., 2025; a Google DeepMind and INSAIT collaboration) filmed 396 real videos of physical scenarios — fluids, collisions, gravity, magnetism, optics — showed models the beginning, and scored their continuations against what actually happened, with the ceiling set by how much two real recordings of the same scene differ from each other. Every model tested, Sora included, scored a small fraction of that ceiling. The pointed finding: the models producing the most photorealistic video were not the ones with the best physics — visual realism and physical understanding were essentially uncorrelated.

Scaling produces case-based mimicry, not laws. A ByteDance study (Kang et al., 2025) built a fully controlled 2D mechanics testbed — balls moving, colliding, bouncing under exact classical mechanics — so that adherence to physical law could be measured, not eyeballed. Video diffusion models scaled from 30 thousand to 3 million training examples generalized almost perfectly within the training distribution, improved measurably on new combinations of familiar elements, and failed outright on out-of-distribution scenarios — with more data and parameters not helping. The failure mode is diagnostic: the models behave as if retrieving the closest memorized training case (matching on color first, then size, then velocity, then shape) rather than applying an abstracted rule. That is Dziri's subgraph-matching result wearing a physics costume, which should give pause to anyone assuming the pathology is specific to language.

Good prediction can hide an incoherent model. Two studies by Vafa and colleagues make this point with elegant directness. In the first (Vafa et al., 2024), a transformer trained on millions of turn-by-turn taxi routes through Manhattan achieved near-perfect next-turn prediction — yet when the researchers reconstructed the street map implied by its predictions, it contained physically impossible geometry, streets that don't exist, and flyovers across Central Park; introduce a few detours and its navigation crumbled. In the second (Vafa et al., 2025), a transformer trained on orbital trajectories predicted planetary motion accurately — but probing what it had internalized revealed no inductive bias toward Newtonian mechanics, and fitting a symbolic expression to its implied forces recovered a nonsensical force law. The model was Kepler without any prospect of becoming Newton: excellent at the sequence, wrong about the world.

The fair counterpoint is Othello-GPT (Li et al., 2023): a transformer trained only on move sequences was shown — via probes that can be manipulated to change the model's behavior — to maintain an internal representation of the board state. So sequence prediction demonstrably can induce structured internal state; the open question is whether it induces state that is coherent, causal, and robust enough to plan with, and the taxi and orbital results say: routinely, no. A useful synthesis of these critiques — and of what a world model would need to be to deserve the name — is the 2025 position paper Critique of World Model.

The verdict on the premise

Grading the original claim piece by piece:

  • "LLMs fail at multi-step reasoning beyond a certain number of steps." Supported, with the refinement that novelty matters as much as length: chains assembled from familiar patterns can run long; chains requiring genuinely novel composition break early. The literature calls this compounding errors or accuracy collapse — model collapse is a different phenomenon (synthetic training data degrading the learned distribution).
  • "World models learn generative properties of the physical world." This is the aspiration, and it is exactly the right description of the research goal. As a description of current systems it is not yet earned: controlled studies consistently find case-based generalization rather than learned physical law, and visually stunning output uncorrelated with physical understanding.
  • "Therefore they can make logical predictions across longer horizons." Partly right, but the mechanism is different from what the phrasing suggests. World models do not escape error accumulation — their rollouts drift too. Where they genuinely extend horizons is in control: abstraction shrinks per-step error, and closed-loop replanning means long-horizon competence never depends on one long fragile prediction. That advantage is proven in games and increasingly in robotics (DreamerV3, MuZero, V-JEPA 2), is backed by theory saying general long-horizon agents can't exist without an accurate world model — and has not yet been demonstrated as a general-purpose reasoning substrate that displaces LLMs.

My own reading: the deepest result in this literature is the Richens–Everitt theorem, and it cuts both ways. Because general agents provably need world models, the research program is well-aimed. And because the theorem doesn't care how the model was learned, LLMs and world models are less rivals than the discourse implies — a sufficiently general LLM agent would have to contain one, and today's explicit world models are, on the evidence, still learning appearances more than laws. The bet embodied by JEPA, Dreamer, and their descendants is that dynamics-first training objectives will get to real physical structure faster than next-token prediction gets there by accident. It is a serious, falsifiable bet. It has not been settled.

What to watch

If you want to track whether the bet is paying off, watch measurements rather than demos:

  • Out-of-distribution physics benchmarks — Physics-IQ scores and successors, especially for models trained with latent-prediction objectives (the JEPA family) versus pixel-generation objectives. The JEPA claim predicts a gap should open; so far the head-to-head evidence is thin.
  • Probing studies in the style of Vafa et al. — does any large-scale model, probed for its implied map or force law, start producing coherent ones?
  • Robots in unfamiliar settings — zero-shot manipulation and navigation results (V-JEPA 2's Franka-arm demonstrations are the current watermark), where closed-loop planning against a learned model either works in the physical world or visibly doesn't.
  • Horizon stability of interactive simulators — Genie 3 holds coherence for minutes; the trajectory of that number is a direct read on the drift problem.

The 2026 investment wave — LeCun's venture, World Labs, DeepMind's and NVIDIA's programs — guarantees the experiments will be run at scale. The honest current summary is that world models are the most credible research direction for AI that deals with physical reality, resting on real theory and real control-domain wins, while the specific claim that they already understand physics fails most careful tests people have devised so far. Both halves of that sentence are worth holding onto at once.

Sources and further reading

Limits of LLM reasoning

World models: foundations and systems

Theory

Do generative models learn physics?

Trying to work out which AI capabilities are real enough to build on — and which are still research bets? Separating the two is a core part of what I do, grounded in a decade of research and industry work.

Get in touch
All posts