PAWBench: How Far Are We from Probabilistically Aligned World Modeling?

PAWBench evaluates video generators as stochastic samplers of world dynamics by measuring probabilistic alignment with actual physical distributions. The benchmark uses PAWEval to assess whether repeated video rollouts recover correct outcome distributions rather than just individual video plausibility. This shifts evaluation from single-trajectory realism to distributional accuracy.