New ViGoR benchmark reveals that leading visual AI models struggle with logical reasoning tasks despite impressive image quality

ViGoR-Bench introduces a unified evaluation framework to expose reasoning deficiencies in modern visual generative models, moving beyond superficial quality metrics

The benchmark uses four key innovations: cross-modal coverage for image and video tasks, dual-track evaluation of processes and results, evidence-grounded automated judging, and granular cognitive dimension analysis

Testing over 20 leading models found that even state-of-the-art systems fail at tasks requiring physical, causal, and complex spatial reasoning

Current evaluations create a 'performance mirage' by overlooking the actual generative process and relying on fragmented benchmarks