Apple research examines LLMs spatial understanding, annotation
…However, the study notes that "Across all models, global conditional counting emerges as a key bottleneck, revealing persistent limitations in compositional and logical reasoning." In other words, most current MLLMs…