Xing Bo's Team Releases WR-Arena Benchmark: No Mainstream World Model Exceeds 60% Accuracy in Environmental Simulation

Recently, the team of Xing Bo, president of MBZUAI (United Arab Emirates AI University) and professor at CMU, released the World Reasoning Arena (WR-Arena). It evaluates mainstream world models—including NVIDIA Cosmos 1/2, Meta V-JEPA2, KLING, MiniMax, Gen-3, and MBZUAI's self-developed PAN—across three dimensions: action simulation fidelity, long-term prediction, and simulation reasoning and planning.

The results show that no model exceeded 60% accuracy in environmental simulation, with a significant gap from human levels. Generation consistency drops sharply with action steps: WAN2.1 fell from 90% to 30% in a 9-step sequence, indicating current models struggle to maintain coherent physical world simulation in multi-step interactions.

MBZUAI's PAN led in most dimensions, though the team also developed the benchmark. The paper concludes that generating realistic videos does not equal understanding how the world works; current world models are still far from being usable as simulators for Agent decision-making.