Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Analogical Reasoning Abilities of MLLMs, Children and Adults on Bongard Problems

Marit Metz¹, Claire E Stevenson¹, Han L. J. van der Maas¹; ¹University of Amsterdam

Presenter: Marit Metz

Bongard problems are visual concept learning tasks that require abstract reasoning and the generation of novel visual concepts (Bongard, 1968, 1970). Despite advances in multimodal large language models (MLLMs), these problems remain challenging for AI. We compare human and MLLM performance using a child-friendly classification version of Bongard Problems. Adults (N=111) outperform all tested MLLMs. Children (N=156, 4-9 years, M=6.19, SD=1.57) perform below adults, but better than Claude and Pixtral and at a similar level as GPT and Gemini models. This is the first study comparing humans and state-of-the-art AI models on a child-friendly format of the Bongard Problems. This study sheds new light on MLLM's (in-)abilities to abstract and generalize in the visual domain compared to adults and children.

Topic Area: Development, Individual Differences & Clinical Populations