Top Neural Networks Failed the ZNO

A team of Ukrainian researchers developed ZNOVision — the first multimodal test for ZNO for AI models. It includes tasks from 13 school subjects, including assignments with pictures and diagrams. The test checks not only knowledge of the Ukrainian language but also understanding of visual context.

No model surpassed the threshold of 70%. The best result was shown by Gemini Pro — 67.5%. GPT-4o — only 47%.

AI performed particularly poorly on visual questions — failing to recognize Ukrainian text in images, ignoring units of measurement, and losing part of the conditions.

Tokensales | News | WaitingRoom#Ukraine