A convincing video clip can still fail as an interactive world: it may ignore a later command, lose track of an object, or return to a changed scene. This benchmark shifts attention from isolated visual quality to whether a generated world remains controllable and coherent over a sequence of interactions.
What It Measures
The 280 cases are divided across four tests: 60 for multi-turn text control, 100 for camera navigation, 60 for interleaved text and camera control, and 60 for spatial memory. Together, they probe different failure modes rather than treating interaction as a single score:
- Instruction following over time: successive prompts test whether actions and subjects remain consistent across generated segments.
- Camera control: navigation cases use 49 scenes and varied camera actions, including paired outbound and return movements.
- Memory on revisits: first-person routes return to distinctive static targets, making changes in scene or object consistency easier to expose.
The release includes structured scene descriptions, interaction sequences, and reference camera trajectories, enabling reproducible input conditions. It supplies evaluation cases and geometry, not a complete scoring system: video comparison and scoring must be handled separately.
Who It Suits—and the Trade-off
A good fit for researchers comparing interactive video world models on controllability, camera response, and persistent scene consistency. The separate task breakdown helps identify which capability is failing, rather than hiding weaknesses in an aggregate result. It is less suitable as a general measure of video quality or broad world-model competence: its scenarios cover a defined set of interactions, and evaluators must choose and report their own scoring procedure.
