Shows that VLMs can overcome deep-seated perceptual biases and optical illusions by using image manipulation tools rather than more training data.
April 1, 2026
Original Paper
Seeing the Evidence, Missing the Answer: Tool-Guided Vision-Language Models on Visual Illusions
arXiv · 2603.29428
The Takeaway
This work shifts the focus from 'fixing' model perception through scaling to using a tool-guided inference framework. It demonstrates that models can arrive at the logically correct answer to an illusion by interacting with the pixels (e.g., drawing lines or cropping regions), decoupling raw perception from reasoning.
From the abstract
Vision-language models (VLMs) exhibit a systematic bias when confronted with classic optical illusions: they overwhelmingly predict the illusion as "real" regardless of whether the image has been counterfactually modified. We present a tool-guided inference framework for the DataCV 2026 Challenge (Tasks I and II) that addresses this failure mode without any model training. An off-the-shelf vision-language model is given access to a small set of generic image manipulation tools: line drawing, reg