WorldMemBench: nine video world models fail revisit memory

A new benchmark called WorldMemBench reports that every one of nine representative video world models it tested degrades on repeated revisits to a scene, and that a model's visual retention does not predict whether its world evolves coherently over time.

The paper runs 260 test cases across 23 semantic categories: 200 for static memory, covering human identity and appearance, object appearance and geometry, and environmental content and geometry, and 60 for dynamic memory, covering event progression and state persistence. It drives each model with "controlled camera departures and multi-revisit trajectories," then scores how closely later frames match the original scene and whether prescribed events actually play out.

The headline split is in the results. "Strong static retention does not imply reliable dynamic state evolution, while repeated revisits reveal memory degradation beyond the first return," the authors write. One example from the write-up: HY-WorldPlay leads all seven static metrics but initiates only 31.7% of prescribed events.

The paper reports that "human studies support the alignment between our automatic evaluations and human judgments," the sanity check new benchmarks usually need and often fail. The authors frame WorldMemBench as "a systematic testbed for diagnosing memory failures and guiding the development of persistent, temporally coherent video worlds." The abstract does not name which of the nine models performed worst overall, nor whether any closed frontier systems were included.