Reasoning models and chain of thought

In short: Reasoning models write a longer chain of thought before answering, working through intermediate steps. This improves results on maths, code and logic noticeably but costs more tokens and time.

In more detail: As early as 2022 “chain of thought” showed that step-by-step thinking raises accuracy. Today it is trained by reinforcement learning: the model gets tasks with a verifiable result (maths problems, tests for code) and is rewarded for correct final results.

In detail

• Why it helps: each generated token is extra compute. Whoever writes more intermediate steps can solve harder problems than in a single forward computation.
• Test-time compute: compute at use time rather than only in training. A “thinking budget” can be set.
• Training: reward for correct answers (RLVR, verifiable rewards). The model discovers strategies by itself such as double-checking, backtracking and starting again.
• Methods for use: self-consistency (several solutions, majority), tree search, verifier models.
• Visibility: some providers show the thoughts, others only a summary.
• Caution: the displayed reasoning is not always the model’s true computation path.
• Costs: many reasoning tokens raise price and waiting time. Simple questions do not need it.

See also: Reinforcement learning, Scaling laws