What's this about?
Video generation models are increasingly described as world models. PAWBench points out that many physical processes can unfold in more than one valid way, so a world model should reproduce not just *a* plausible trajectory but the distribution of possible ones. 136 upvotes.
Why does the distribution matter more than the single outcome?
Drop a glass and it shatters — but into which pieces, in which directions, is genuinely random within a range. A model that always produces the same shattering pattern has memorised an outcome, not learned the physics. Worse, if you use it to plan, it will be confidently wrong about how much variation to expect. Reproducing the distribution is what separates a model that understands a process from one that can render a convincing example of it.
Why does this matter to you?
- It is the same critique VGI-Bench made of video-model reasoning, applied to physics rather than to visual inference.
- Distributional evaluation is much harder than sample evaluation, which is why almost nobody does it — and why benchmarks reward memorisation.
- Directly relevant to anyone treating a generative model's single output as a prediction rather than a sample.