A reasoning model is an LLM optimized for demanding problem solving. It generates intermediate reasoning tokens to break a task into smaller steps before producing a final answer. This extends the same next-token prediction process over a longer chain, increasing token use and response time; reasoning tokens also occupy the context window. section 1.2.1
Prompt these models by specifying the task and its grounding data, constraints, examples, and tool definitions. Explicit procedural directions such as chain-of-thought, tree-of-thought, or self-consistency prompting are largely redundant and can constrain the model’s trained reasoning process. section 6.3.7






