Reasoning
A model often fails not because it cannot solve a problem, but because it commits to the first token of an answer too quickly. Reasoning models are trained to work through steps before answering.
Instead of jumping straight to a conclusion, the model generates a chain of intermediate steps, each token conditioned on the ones before. Spending more tokens on this hidden working is test-time compute: the same model gets more accurate on hard problems simply by thinking longer. The trade is latency for reliability.
A bat and ball cost $1.10. The bat costs $1 more than the ball. How much is the ball?
Check yourself
The fast and slow paths use the same weights. Why does forcing intermediate steps change the final answer?
Go deeper (free): Lilian Weng — Why we think (test-time compute) ↗