Benchmarking Vision-Language Models on Visual Illusion Understanding
How to measure whether a vision-language model perceives visual illusions the way people do: assembling a dataset that mixes classic synthetic illusions with real-scene illusions, color-blindness plates, and trap images; probing it with true-or-false, multiple-choice, and open-ended description tasks against human baselines; and testing prompting strategies that make the model name an illusion's cause before describing content. Students learn why real-world and anti-overfitting images matter and how step-by-step reasoning changes model accuracy.
IllusionBench+: A Large-scale and Comprehensive Benchmark for Visual Illusion Understanding in
A computer-vision benchmark paper introducing IllusionBench+, a large-scale dataset for testing whether vision-language models (VLMs) understand visual illusions. It contains 1,051 images - classic c…