WorldExam: no video world model handles all three control types

A new benchmark out of CASIA and collaborators quietly reframes what a 'world model' should be graded on, and the headline finding is that none of the twenty systems tested clears the bar. The paper, WorldExam on Hugging Face, argues that current evaluation focuses on _apparent appearance_ (does the video look right, does it follow the instruction) and skips _inherent reactivity_: whether the world reacts to a subject the way a real world would when the instruction stops short of spelling it out.

The construction is worth pausing on. WorldExam splits evaluation into four diagnostic levels, Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity, and instantiates them through eight tasks including Camera Control, Subject Control, Scene Revisit, Terrain Interaction, Object Interaction, Social Interaction, Physical Reaction, and Goal Completion. That is 1,474 cases in total, run against 20 representative models across camera-, action-, and language-driven paradigms, with a static-scene track that any interface can attempt and a dynamic-interaction track that only the action- and language-driven models can. First-person viewpoints make up 31.4% of the benchmark and no scene category dominates, with traffic scenes topping out at 14.7%.

The finding is a clean three-way split. Camera-driven models excel at camera control, but their interfaces do not support dynamic interaction. Action-driven models control subjects more precisely but often leave the world unresponsive. Language-driven models perform better on interaction tasks but follow complex controls less faithfully. Put plainly, the paper's own summary is that no model combines broad task coverage with consistently strong performance, and that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity.

The honest caveats matter. The abstract and framing sections do not publish a per-model leaderboard, so which twenty systems win which task is not visible from what I retrieved. Four of the eight tasks lean on GPT-5.5 as a VLM judge scoring against case-specific checklists, which is a reasonable choice but one worth noting for anyone treating the results as ground truth. The authors say they will publicly release the benchmark data and evaluation toolkit; timing and license are not in the text I saw.

Still, the direction is the interesting part. If world models are going to underwrite robotics simulators, autonomous-driving stacks, and interactive gaming, the evaluation had to move past 'looks plausible' toward 'reacts like a world.' WorldExam is a bid to make that shift measurable, and the fact that nobody currently passes is a healthier baseline than a saturated leaderboard would be.