WorldOlympiad: Can Your World Model Survive a Triathlon?
TLDR
WorldOlympiad benchmark evaluates video-based world models across physical, geometric, and interaction fidelity using real-world scenarios.
Reasoning
The paper introduces a novel benchmark that decomposes world-model evaluation into three complementary dimensions, covering gaming, robotics, and real-world videos, which is a strength. However, the abstract is truncated and does not provide full experimental results or limitations, making it difficult to assess the depth of analysis.
Read-first score
Read-first score 66.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 59.
Field roles
Rank sensitivity
Stability: volatile; rank range: 428.
Keyword Scores
Deep Analysis
Innovations
- Decomposes world-model evaluation into three complementary dimensions: physical faithfulness, geometric consistency, and interaction fidelity
- Covers three major downstream scenarios: gaming, robotics, and general real-world videos
- Uses object segmentation and MLLM-as-judge for physical track, Gaussian splatting for geometry track, and action prompts for interaction track
Methodology
WorldOlympiad evaluates video-based world models across three tracks: physical faithfulness (using object segmentation and MLLM-as-judge to assess mechanics, thermal phenomena, and material properties), geometric consistency (reconstructing videos with Gaussian splatting to evaluate structural consistency, cross-view coherence, and camera-trajectory alignment), and interaction fidelity (assessing whether generated rollouts follow complex action prompts and maintain smooth transitions across consecutive video chunks). It covers gaming, robotics, and general real-world scenarios.
Key Results
Experiments on state-of-the-art models reveal substantial gaps in physical reasoning, 3D consistency, and long-horizon interaction, underscoring the need for more structured evaluation protocols for generative world models.