Get the latest tech news
Vision Language Models Are Biased
Vision Language Models are Biased: VLMs fail on simple counting tasks when familiar objects are subtly modified
Mean Accuracy: 0.44% Counting circles in Audi and points in Mercedes star Key Finding: Worst performance in logos category. Small logo size relative to the vehicle made visual bias even stronger - models completely ignored modifications. Models scored 0% on Sudoku and Go boards, confirming fundamental inability to perform basic visual counting in structured settings.
Or read this on Hacker News