Researchers improve vision-language accuracy by fixing SPaRC images

Researchers evaluating vision-language models on spatial grid puzzles found that changing the visual input had a vastly greater impact on accuracy than rewording the prompt. Adding targeted annotations that fixed misread elements and linked the image to task terms more than doubled model scores, whereas adding general text or clutter often reduced accuracy. The author argues that developers should check what a vision model actually sees and adjust the image before tuning prompts. Open article →