PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives

TL;DR - PlayWorld benchmarks video world models using multimodal agents pursuing long-horizon objectives rather than fixed action sequences. Tests across nine models expose persistent weaknesses in spatial consistency and state evolution.

• Provides 171 objective-driven interactive scenarios.
• Evaluates geometry consistency, interaction fidelity, and visible and out-of-sight evolution.
• Also measures basic video quality and action controllability.
• Agent players enable fairer comparisons when different models require different action sequences to reach the same goal.