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

arXiv 2026 43.8 benchmark, theory

TLDR

Introduces PAWBench and PAWEval to evaluate whether video generators reproduce distributions of physical behaviors under identical observations and actions; current systems fail probabilistic alignment.

Reasoning

The paper provides a clear formalization and a large benchmark with 50 scenarios and 11 systems, making a convincing empirical case that current video generators are not probabilistically aligned. However, the abstract only sketches interventions to improve alignment and does not report detailed results, so the practical path forward remains unclear.

Read-first score

Read-first score 43.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=reproduce

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 385.

Keyword Scores

world model

generative world model

video world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Formalizes probabilistic alignment as a distributional criterion for world models.
Introduces PAWBench, a benchmark for evaluating video generators as stochastic samplers of world dynamics.
Introduces PAWEval, an outcome-level protocol that converts repeated video rollouts into empirical distributions over possible physical behaviors.

Methodology

The paper formalizes probabilistic alignment for world models and introduces PAWBench with 50 scenarios. It evaluates eleven current video generation systems using the PAWEval protocol, which converts repeated video rollouts into empirical distributions over possible physical behaviors. It also tests whether language prompts, initial noise sampling, or model training can reshape predictive distributions.

Key Results

Across 50 scenarios and eleven current systems, no model consistently matches the reference probabilities while recovering the range of valid behaviors.

Resources

Paper