Before it
Reasoning was prompted at inference; training on reasoning needed human-written rationales.
The problem
Nobody can write rationales for every problem a model should learn to solve.
What it introduced
Let the model generate its own rationales, keep the ones that reach the right answer, and fine-tune on those.
Why it matters
Self-generated reasoning data, verified by outcome, is the core loop that RL-trained reasoning models later industrialised.
What modern models inherit
Training on a model's own verified reasoning.
What came next
Process supervision, then the RL reasoning training in o1 and DeepSeek-R1.