Video generation has produced a lot of good models recently, but "good" keeps drifting. One paper reports FVD, another claims long-horizon consistency, a third leans on physical plausibility, and three vocabularies sit there with no common ruler between them.
A 102-page survey on arXiv (https://arxiv.org/abs/2610.31668), submitted on 22 October and led by the University of Oxford together with the National University of Singapore, Nanyang Technological University and the University of California, Merced, tries to supply that ruler. Its judgement is blunt: if a piece of work wants to call itself a world model, on this scale it has to reach at least the third rung.
One coordinate system: reading, writing, sharing, interacting
The survey’s move is to reframe what a video is. A video is not the world; it is a partial, lossy, temporally ordered projection of a world that keeps evolving. A car drives behind a wall and the frames show nothing. The car is still in the world.
Under that lens, the field’s separate communities collapse into four operations:
- Reading: infer properties of the world from observed video.
- Writing: synthesize video from specified or learned world conditions.
- Sharing: integrate understanding and generation through common modeling mechanisms.
- Interacting: predict and update world states under actions and evolving observations.
The project page gives each operation its own evaluation criteria, instead of letting understanding and generation keep sharing one scoreboard.
A 2x3 grid for placing models
The authors draw a 2x3 taxonomy of video world models. One axis is prediction space: observation-space models predict frames, video tokens or video-decoding latents, while latent-state models predict video-derived features without requiring rendering. The other axis is transition interface: passive prediction with no action input, action-conditioned prediction with controls fixed in advance, and interactive prediction that accepts new controls after feedback.
The placement of familiar names is the interesting part here. Sora, V-JEPA and Genie 3 each occupy a cell, and none of them is the whole grid. The same model can sit in different cells under different training and use settings.
There is also a qualifier that is easy to skim past: "interactive" describes the control interface, not the attention pattern. Video-decoding latents still belong to observation space.
An evidence ladder: rung three to call it a world model
This is the sharpest part of the survey. Each operation gets an evidence ladder, and the same four rungs recur:
- Plausible output - credible answers, frames or features on the immediate task.
- Faithful conditions - dependence on observed evidence and supplied controls.
- Consistent world state - persistent identity, layout and physics, with controlled transfer for sharing.
- Decision utility - improved planning, policy learning or policy evaluation.
The authors’ formulation is blunt: a "world model" claim needs the third rung. On that standard, a good deal of work that markets itself as a world model is actually stuck on the first.
Seven open challenges, each with a usable test
The most practically useful part of the survey is that it never lists a problem without attaching a way to measure it.
- Long-horizon world consistency: vary rollout length and revisit distance separately, report degradation alongside memory, latency and action-update rate.
- Physical and causal grounding: compare paired interventions from the same initial state, and locate the first violated constraint and when it appears.
- Measurable synergy in unified models: compare coupled and uncoupled systems under matched data, capacity, compute and supervision. Doing both tasks does not establish that understanding and generation improve one another.
- Standardized evaluation of world properties: align state properties and interventions across operations, and report protocol, budget, failures and downstream effects.
- Scaling control from unlabeled video: video shows motion but omits the actions, units and coordinate systems that caused it, so test whether inferred controls actually predict interventions.
- Generalization across scenes and embodiments: separate held-out scenes from held-out embodiments, then evaluate their combinations under stated adaptation budgets.
- Stochastic worlds and rare events: use calibration, proper scoring rules and valid-mode coverage at fixed sample budgets, and count invalid outcomes separately.
The benchmark examples show how concrete this gets. RoboWM-Bench converts generated manipulation videos into actions and executes them in a physics simulator, so visually plausible but unexecutable motion earns little credit. EWMBench separates scene consistency, motion correctness and semantic alignment instead of collapsing them into one score.
Why the current metrics fall short
The survey also dissects the metrics the field runs on. FVD depends on the feature extractor and the sample count, and can under-penalize temporal corruption in favor of static samples. Inception Score rewards confident and diverse classifier predictions without any reference to real data. PSNR, SSIM and LPIPS suit reconstruction and source preservation, but they penalize valid alternative futures: a model that draws a different but reasonable trajectory loses points for being right.
The deeper problem sits on the reading side. Does the answer actually depend on the video at all? The survey proposes a set of cheap controls: text-only, single-frame, shuffled-order and restricted-context. A model that fails all four is not using the video.
So what
The scarce value of this survey is not the summary, since anyone can summarize. It is that it sets an entry bar for the term "world model", and lays the blind spots of existing evaluation metrics on the table while doing it.
For people building models: before claiming a world model, work out which rung your evidence actually reaches. For people building evaluations: the next useful step is not another distributional distance, but making intervention-style tests the default, because without intervention you can never check whether the world is right.
For you: the project you have been calling a world model, which rung is it on?