Benchmark RadarPlayWorld
PlayWorld evaluates interactive video world models across 171 scenarios with long-horizon objectives. Agent players adaptively control each model, and performance is scored on geometry consistency, interaction fidelity, out-of-sight evolution, and insight evolution, plus basic video quality and controllability metrics.
- Released
- 2026-08-13
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Standard fixed-action evaluations fail to compare world models that handle actions differently. PlayWorld provides an objective-driven protocol that reflects real user interaction, enabling fair cross-model assessment and highlighting limitations in spatial consistency and state evolution for long-horizon tasks.
Motivation
Video world models simulate future states conditioned on current observations and user actions.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.