Visual Puzzles that Still Make AI Go Mad?
In 1967, Mikhail Bongard introduced a set of visual pattern recognition problems that would become a cornerstone benchmark for artificial intelligence. A Bongard problem presents two sets of images; the AI must find the rule that distinguishes Set A from Set B. Simple for humans. Infuriatingly difficult for machines.
While modern deep learning models have conquered many benchmarks, Bongard problems remain a fortress. They require a uniquely human ability: the capacity to infer abstract rules from just a handful of examples, often relying on spatial relationships, counting, or shape transformations that neural networks struggle to generalize.
In our research, we decided to push these AI systems into the absurd. By generating increasingly surreal and complex Bongard problems—much like the shifting logic of Alice's Wonderland—we expose the brittle nature of current visual reasoning architectures.
We found that even the most advanced Vision-Language Models (VLMs) and Large Language Models (LLMs) fall apart when the visual logic steps outside their training distribution. They hallucinate rules that aren't there, and confidently fail to see patterns that are obvious to a child.
Now it's your turn. Can you solve what the AI cannot?
Click the button below to begin the interactive Bongard evaluation.
The "Bongard in Wonderland" project highlights a fundamental gap in machine cognition. While AI excels at pattern matching within massive datasets, it lacks the compositional generalization that human reasoning employs effortlessly.
Understanding this limitation is the first step toward building truly hybrid systems—where human intuition guides machine computation. This is the core mission of the TANGO project.