The model thinks before it answers
The idea: Reasoning models are trained to spend tokens thinking first — a private chain of thought — before the final answer, so the 'scratch paper' is just more generated tokens.
What you'll be able to do: You can explain why reasoning models think before answering, and that the 'thinking' is generated tokens.
The problem it solves: Hard problems need working-out. But the model answers in a single pass — no scratch paper to think on.
Builds on: Decoding & sampling knobs
← GPT-2 → GPT-3 → frontier · Next: Test-time compute: pay at answer-time →
All lessons