Reasoning model
Beginner
An AI that writes out its working before it answers, the way you might on a hard math problem.
Novice
A language model trained or prompted to produce intermediate steps (its “thinking”) before the final answer. It spends more computation on each question in exchange for better answers on multi-step problems.
Expert
An LLM whose post-training (typically RL with verifiable rewards, SFT on long traces, or distillation) makes it emit long chains of thought, often between delimiters such as think tags, before answering. Accuracy scales with the number of thinking tokens, at a matching cost in decode time and KV-cache memory.