How Far Has AI Come in Abstract Visual Reasoning? Insights from Bongard Problems

Abstract Visual Reasoning (AVR) encompasses a diverse range of tasks that require discovering patterns, inferring relations, and identifying abstract concepts from visual information. Such tasks provide a valuable testbed for assessing reasoning capabilities in both humans and AI systems.

In this talk, I will first introduce the major types of AVR problems and outline the main approaches to solving them, highlighting how the performance of contemporary AI varies across different task categories. I will then focus on Bongard Problems (BPs), which require discovering and articulating an abstract concept that distinguishes two sets of images. BPs pose a particularly interesting challenge for vision-language models (VLMs), as solving them requires integrating visual perception, abstract reasoning, and natural language.

Drawing on several years of research, I will trace the progress of successive generations of VLMs in solving synthetic and real-world Bongard Problems. The experimental results will illustrate substantial improvements in model performance, while also revealing persistent limitations. Finally, I will reflect on what these findings tell us about the current capabilities of AI in abstract visual reasoning and the challenges that remain on the path toward more general intelligence.

Jacek Mańdziuk is a Full Professor at Warsaw University of Technology (WUT), Poland, and an elected Corresponding Member of the Polish Academy of Sciences. At WUT, he serves as Director of the Center of Excellence for Artificial Intelligence and Head of the Division of Artificial Intelligence and Computational Methods. He is also Co-Director of ELLIS Unit Warsaw, a Member of the Council of Scientific Excellence (Poland), and an IEEE Senior Member.

His research interests include artificial general intelligence, with particular emphasis on cognitively inspired learning and reasoning, abstract visual reasoning, and human-like problem solving, as well as evolutionary computation, metaheuristic optimization, and AI for social good. He has authored or co-authored over 250 scientific publications. His work has appeared at leading AI and machine learning conferences, including ICML, ICLR, AAAI, and IJCAI. He has also delivered more than 40 invited lectures at international academic institutions.

He was a Fulbright Scholar at the University of California, Berkeley, and at the International Computer Science Institute (ICSI) in Berkeley, USA. He was also a Visiting Professor at Nanyang Technological University, Singapore. He served as General Co-Chair of the IEEE Congress on Evolutionary Computation (CEC 2021) and chaired the annual IEEE SSCI Symposium on Computational Intelligence for Human-like Intelligence from 2013 to 2023.