Researchers from several US universities and the startup Cursor have developed a new test to assess the capabilities of generative AI models. They used puzzles from the radio show “Sunday Puzzle,” broadcast on NPR. This test revealed unexpected features in model behavior, such as the fact that some models, like those from OpenAI, sometimes “give up” and provide incorrect answers.
Interestingly, the test includes puzzles that are understandable without specialized knowledge, making it accessible to a wide audience. The “Sunday Puzzle” does not require models to have specific expertise, and the problems are formulated so that models cannot rely on “mechanical memory.” This makes the test appealing to researchers seeking to understand how AI models solve tasks that require intuition and the process of elimination.
Currently, the best results on the test were achieved by the o1 model with a score of 59%, while the new o3-mini model, tuned for high reasoning effort, scored 47%. The researchers plan to expand testing to other models to determine how their performance can be improved. This could help identify which aspects of model operation need enhancement.
However, the “Sunday Puzzle” test has its limitations, as it is aimed at an English-speaking audience. Nevertheless, the researchers believe that regularly updating the questions will help keep the test relevant and allow them to track how model performance changes over time.