How Little Do World Models Understand the World? Xing Bo's Team Releases WR-Arena Benchmark, No Model Exceeds 60% in Environmental Simulation

According to monitoring by 1M AI News, Xing Bo’s team from the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and a professor at Carnegie Mellon University has released the World Reasoning Arena (WR-Arena), which systematically evaluates current mainstream world models from three dimensions: action simulation fidelity, long-term prediction, and simulated reasoning and planning. The test subjects include NVIDIA’s Cosmos 1/2, Meta’s V-JEPA 2, KLING, MiniMax, Gen-3, and MBZUAI’s self-developed PAN. The results are not optimistic: no model exceeded 60% accuracy in the environmental simulation dimension, indicating a significant gap compared to human levels. Generation consistency sharply declines with the number of action steps; WAN 2.1 dropped from 90% to 30% in a 9-step action sequence, meaning current models are nearly incapable of maintaining coherent simulations of the physical world in multi-step interactive scenarios. MBZUAI’s self-developed PAN leads in most dimensions (transition smoothness 53.6%, generation consistency 64.1%, open planning improvement 26.7%), but the PAN team is also the benchmark’s creator. The core conclusion of the paper is that being able to generate realistic videos does not equate to understanding how the world operates; current world models are still far from being usable simulators for agent decision-making.