Scaling test-time compute paper argues extra inference compute can beat bigger models
On some problems, extra inference-time computation matched the gains from a pretrained model roughly 14 times larger, the authors reported.
Ideas & essays · Benchmarks & progress