WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models
Video models that retain a scene’s appearance can still fail to keep its events and states consistent.
WorldMemBench tests nine models across 260 cases covering people, objects, environments, event progression and state persistence. Controlled camera departures and repeated revisits check whether content survives and events follow expected state changes. Repeated returns expose memory degradation beyond the first revisit. Human studies support the automatic evaluations’ alignment with human judgments.
HF Daily Papers' note
WorldMemBench tests nine models across 260 cases covering people, objects, environments, event progression and state persistence. Controlled camera departures and repeated revisits check whether content survives and events follow expected state changes. Repeated returns expose memory degradation beyond the first revisit. Human studies support the automatic evaluations’ alignment with human judgments.
HF Daily Papers' note
score 5