What should post-training optimize? A test-time scaling law perspective

Muheng Li, Jian Qian, Wenlong Mou

Large language models are increasingly deployed with test-time strategies: sample $N$ responses, score them with a reward model or verifier, and return the best. This deployment ru…