The one-sentence takeaway

Grading every step of a model's reasoning, rather than only its final answer, produces a verifier that is much harder to fool, picks correct solutions to hard problems far more reliably, and rewards genuinely sound thinking instead of lucky guesses.

Next: Chapter 4, DeepSeek-R1, where a model learns to reason not from human examples but from reinforcement learning.