Problems require a numerical answer and often involve algebra, geometry, number theory, or combinatorics. Model evaluations typically use released contest questions and exact answer checking under a stated prompting setup. Training-data exposure and repeated tuning can affect how the score should be interpreted.