Abstract

Humans don’t learn intelligence—they develop it. Biological life on Earth benefits from innate cognitive primitives encoded in its genes. Cats don’t learn to backflip, just as birds don’t learn to fly. Likewise, humans don’t learn to think; rather, we are born with core cognitive abilities that shape our perception and interaction with the physical world. These foundational cognitive structures unfold along a developmental trajectory as we mature, which has been studied extensively by cognitive developmental psychologists in the past hundred years.

But do machines require the same core knowledge to achieve human-like general intelligence? To investigate the answer to this question, we conduct a series of empirical and theoretical studies to explore the extent to which current Multimodal Large Language Models (MLLMs) capture core knowledge. We have curated a large-scale, cognitive science-inspired benchmark spanning 12 core cognitive concepts and four Piagetian developmental stages. Evaluating 219 models across 10 different prompts, we generated 2,409 data points for analysis. The results reveal significant deficits in early-developed core abilities, even as models achieve human-comparable performance in other general tasks. Notably, fundamental core cognitive abilities do not benefit from scaling. We also examine the philosophical foundations of growing AI like a child, considering the insights from the past 100 years of human cognition research.

Bio

Bio of Tianwei Zhao:  Tianwei Zhao is a master’s graduate in Computer Science from Johns Hopkins University. His research focuses on evaluating, interpreting, and enhancing reasoning and planning in multimodal large language models from a cognitive science perspective.

Bio of Hokin Deng: Hokin Deng is an independent researcher based in New York, United States. Before that, he was trained in neuroscience and philosophy from Johns Hopkins University.