Benchmark RadarWorldOlympiad
WorldOlympiad is a triathlon-style benchmark for video-based world models, evaluating physical faithfulness, geometric consistency, and interaction fidelity across gaming, robotics, and general real-world scenarios. It includes 1,000 long videos and interpretable automatic metrics.
- Released
- 2026-06-09
- Readiness
- Runnable
- Primary field
- Robotics & Autonomous Systems
Why it matters
Existing video generation benchmarks focus on visual quality or short-term coherence, lacking diagnostics for physical rule compliance, 3D structural consistency, and long-horizon interactive control. WorldOlympiad provides a structured protocol to identify specific failure modes in world models, informing model development and selection for embodied and interactive applications.
Motivation
We introduce WorldOlympiad, a benchmark for diagnosing video-based world models across physical faithfulness, geometric consistency, and interaction fidelity.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.