“A key strategy for improving perceived model performance is "test-time compute," which involves allocating more computational resources during inference to generate and evaluate multiple potential answers.”