A stupid question: Why researchers call such methods as "scaling test-time computation"?

[huanranchen]

Thanks for your great work!
I'm curious about why researchers refer to such methods as "scaling test-time computation." From your blog, it seems there isn't an explicit scaling of test-time computation. It appears that models trained with simple PPO tend to generate longer answers.

[Benjoyo]

Because the models produce better results by generating a lot of „ thinking“ tokens compared to answering ad-hoc. Hence: more inference compute traded for better accuracy