Five failures that outlived every fix: persistence, action, horizon, target, and verification. Why the measurements arrived in the wrong order, and four questions every paper should answer.
Five failures, five communities, and why the most important ones were measured last.
Every system in this course ends on something that does not work yet. Five failures kept turning up: persistence, action, horizon, target and verification. Each was found by a different community, and each was measured late or not at all. The confusion around the phrase “world model” was built out of them.
A rollout is the model’s predictions run forward step after step, each one built on the last. The usual report is one average error over the whole run. That melts several failures into one number and hides the order they happened in. Identity can go while texture stays sharp. Action fidelity can go while both still look fine. The useful report is one horizon per property: how many steps each one held up for.
Object permanence is the expectation that a thing is still there, and still itself, when you look back. It is the failure everyone notices. A renderer has no object store, so identity lives in the predictor’s hidden state. When the chair becomes another chair, it does so a little at a time and without an error message. I call this the single-frame alibi: you get a rollout a hundred steps past the point of being useful, and any one frame of it still looks completely fine. The mistake only shows when you look along the sequence.
Colour and texture outlast identity and physics, and the loss explains why. A per-frame loss pays the model for getting most of the pixels about right, and most pixels are surfaces and light. Nothing in it is watching whether the chair you walked past is the same chair.
A model can tell you what is likely to happen next. What would have happened if you had done something else is a different question, and it is the one a decision needs. The model’s answer to it comes from the model, not from the world.
Bingyi Kang and colleagues measured the gap in 2024, in How Far is Video Generation from World Model. They tested video generators on physical laws inside and outside their training data. Inside, the generators obeyed the laws. Outside, they broke them — and nothing in the output said which case you were in. A system that said “I have never seen anything like that” would be more useful and less impressive, and nobody is rewarded for building it.
Most tricks in model-based control are ways of not needing the model to be right for very long. Horizon was the first of the five to be measured — measured by the people who had to act on the model’s predictions, in reinforcement learning.
Tingwu Wang and colleagues did it in 2019 in Benchmarking Model-Based Reinforcement Learning. The planning horizon — how many steps ahead the model was trusted — decided where each method won and lost. Michael Janner and colleagues answered the same year, in When to Trust Your Model: only briefly, and from a real state. Short rollouts and replanning every few steps work because they stop relying on the model before it drifts.
A representation, an embedding learned for something else to use, can look stable while throwing away the variable the task needed. I call that the target problem: predicting the wrong thing well. The credit assignment problem in representation learning — which details to keep and which to discard — is not solved by training well. It depends on what the downstream task needs, which the representation does not know.
Quentin Garrido and colleagues showed in their 2024 paper Learning and Leveraging World Models in Visual Representation Learning that the strength of the predictor decides what the encoder keeps. Give the predictor an easy job and the encoder throws away whatever the transformation changed.
Measurement arrived in the wrong order. Reinforcement learning measured horizon first, because an agent that trusts a bad model loses, and the loss shows up in the score. The video people came next and measured what they could, which was the frame. VBench came in 2023, from Ziqi Huang and colleagues, scoring generated video on separate dimensions. PlayWorld, released in August 2026, tests 171 scenarios covering geometry, interaction fidelity and object persistence — the two failures a renderer hits first, tested seven years after horizon.
Frame quality is measurable and it is what demos are made of. Action fidelity separates a world model from a video generator, and it is almost never reported. That order is close to the reverse of how much each matters.
An implicit model is a claim that a model formed inside a network trained for some other job. Checking it needs a probe — a small second network trained to read one fact out of the first’s internals. Kenneth Li and colleagues showed what that takes in 2022, in Emergent World Representations. A network trained only to predict the next move in Othello turned out, under probes, to hold a picture of the board. The probe could have found nothing. The four questions this raises: the probe has to be small (a big probe can extract structure from a random network), the feature has to be tested on held-out positions, the probe accuracy has to beat chance by a meaningful margin, and the internals have to be causally upstream of the behaviour you care about.
Four questions handle most papers.
The first is what does it output? A picture, a structure, a compact state or an embedding — and the other three questions depend on the answer.
The second is can you roll it forward under actions nobody took? If not, it will not support a decision, however good it looks.
The third is over what horizon does it hold up, and who measured that? The number matters more than the architecture, and it is usually missing.
The fourth is which measure is the claim using, and what does it ignore? There is no overall score, so anyone implying one has picked a measure and not told you which.
Yann LeCun’s 2022 paper A Path Towards Autonomous Machine Intelligence sets out what a world model would have to do to deserve the name. Read against the four questions, it is a list of what is still missing, and naming an architecture answers none of them.
Say the answer out loud before you reveal it — recalling it is what makes it stick, and rereading it is not.
What is the single-frame alibi, and which failure does it describe?
Why is object permanence hard for learned world models?
Why was horizon the first failure to be measured, and what did Janner et al. recommend?
What are the four questions every world model paper should answer?
5 failures: Persistence (single-frame alibi — chair drifts unseen), Action (counterfactuals fail outside training band; Kang 2024), Horizon (errors compound; trust only briefly — Janner 2019), Target (predict wrong thing well; downstream task decides), Verification (measurements in reverse of importance). 4 questions: output type? counterfactual support? measured horizon? which measure? LeCun 2022: naming an architecture answers none of them.