WorldReasonBench exposes poor logic in top AI video generators
a tsinghua university benchmark shows commercial ai video models like seedance 2.0 outperform open-source rivals, yet all fail at logical and physical reasoning tasks.
worldreasonbench evaluates models like sora 2 and veo 3.1 on their ability to continue starting scenes with physical, social, and logical plausibility, finding that high visual fidelity often masks fundamental errors in cause and effect. commercial systems scored double the open-source models on reasoning metrics, with seedance 2.0 taking the top spot, though logical reasoning remained the weakest category for everyone. open-source models showed greater dependence on detailed prompts to compensate for their lack of built-in reasoning capabilities compared to commercial rivals. the benchmark uses a two-stage scoring system, validated against 6,000 human annotations, to prove that current models have not yet become reliable world models.