Can AI models keep one scene in mind? Models find the one view of a room that contradicts the other three.
Source: the SSoN paper (Table 2). Six vision-language models, evaluated without camera information or 3D reconstruction.
Each exam is built on a question that scientists are still working to answer.
What an agent produces is checked the way a field checks new results, against data and its established standards.
Each exam starts from what a field already knows and asks what an agent’s approach can add to it.
AI's Exam is a collection of exams for AI, starting with climate science and vision. Each topic poses its own problem with its own data, tasks and evaluation, and each has its own page with the full results.