cs.CVAug 27, 2026

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

Authors: Yuandong PuLe ZhuoSayak PaulGabriel Jorge MenezesAvram ĐorđevićShiyang LiYifan ZhouBin Fu+6 more

Organizations: Shanghai Jiao Tong University · Shanghai AI Laboratory · Krea AI · Hugging Face · Shanghai Innovation Institute · Tongyi Lab · The University of Hong Kong

Abstract

Recent video generation models are increasingly framed as world models. Many physical processes can unfold in more than one valid way. Therefore, a world model should reproduce not only a plausible trajectory, but also the distribution of possible behaviors under the same initial observation and action. We call this distribution-level requirement probabilistic alignment. However, existing evaluations largely assess individual-video plausibility and do not test whether repeated generations recover the correct distribution. This raises a central question: how far are current video generators from probabilistically aligned world modeling? To answer it, we formalize probabilistic alignment as a distributional criterion for world models and introduce PAWBench, a benchmark for evaluating video generators as stochastic samplers of world dynamics. We further introduce PAWEval, an outcome-level protocol that converts repeated video rollouts into empirical distributions over possible physical behaviors. Across 50 scenarios and eleven current systems, no model consistently matches the reference probabilities while recovering the range of valid behaviors. Having established this gap, we test whether language prompts, initial noise sampling, or model training can reshape the model's predictive distribution. We believe our work can serve as a foundation for future efforts to move towards probabilistically aligned world modeling.

Explore similar work

CardsList
  1. HappyWorld-Bench

    Sep 21, 2026Zhiqi Bai, Junai Cai, Yixin Chen +33