“According to research by Snell et al. (2024), easier questions benefit from purely sequential test-time compute, whereas harder questions perform best with an optimal ratio of sequential to parallel compute.”