PAWBench finds no world model matches 50 scenes

PAWBench tested 11 video-generation systems across 50 physical scenarios and found that none consistently matched the reference distribution of valid outcomes, according to the benchmark paper. It asks whether a purported world model can reproduce multiple plausible futures—not merely one convincing clip—so the result is a warning against treating visual plausibility as robot-planning evidence; teams should add repeated-rollout distribution tests beside the archive’s world-model simulation analysis before letting synthetic footage set a control policy.