Machine Learning Research

Multimodal large language models exhibit systematic failures in spatial reasoning and temporal understanding, with cascading errors emerging from initial perception mistakes. Clock-reading tasks—requiring identification of hour and minute hands plus spatial positioning—reveal critical gaps that propagate through subsequent analysis steps. Enterprise adoption faces reliability barriers as models struggle with variations that humans process effortlessly.