PAWBench: How Far Are We from Probabilistically Aligned World Modeling?
The benchmark says current video generators still fail at matching the probabilities of possible physical outcomes.
PAWBench tests whether repeated video rollouts recover the right distribution of valid behaviors, not just whether one generated clip looks plausible. The authors pair it with PAWEval, which turns those rollouts into empirical outcome distributions across 50 scenarios. Across eleven current systems, no model consistently matched the reference probabilities while also covering the range of valid behaviors. The paper also probes whether prompts, noise sampling, or training can shift those predictive distributions. HF Daily Papers' note
PAWBench tests whether repeated video rollouts recover the right distribution of valid behaviors, not just whether one generated clip looks plausible. The authors pair it with PAWEval, which turns those rollouts into empirical outcome distributions across 50 scenarios. Across eleven current systems, no model consistently matched the reference probabilities while also covering the range of valid behaviors. The paper also probes whether prompts, noise sampling, or training can shift those predictive distributions. HF Daily Papers' note
score 5