[HUGGINGFACE]score: 0.63
CRONOS Benchmark Tests Whether Video Models Learn Causal Physics or Visual Shortcuts
May 21, 2026
CRONOS uses photorealistic Unreal Engine environments to apply controlled interventions on viewpoint, scene context, object appearance, and category, measuring whether video prediction models respond appropriately to counterfactual physical changes rather than exploiting visual correlations.
paper
HOW THIS AFFECTS YOU
●
researcherCRONOS provides a systematic intervention-based evaluation that can distinguish causal physical understanding from pattern matching in video world models, addressing a key gap in existing benchmarks.