Skip to content
Search
TechniquesGlossary

Reasoning Model

Classic chain-of-thought prompting asked a model to think step by step. Reasoning models are trained to do this on their own, usually with reinforcement learning that rewards correct final answers. DeepSeek-R1 is one openly documented example: its authors report that reinforcement learning alone, without human-written reasoning examples, produced reasoning behaviors such as self-reflection and verification. Most major model providers now offer reasoning models or a reasoning mode, often with a setting that controls how much thinking to allow.

Prompting them is a little different. Telling a reasoning model to think step by step is usually redundant, and long worked examples can get in its way. State the goal, the constraints and what a correct answer looks like, then let it work. Some providers restrict sampling settings such as temperature on these models, and the thinking counts toward your output token limit, so check the API reference for the model you use. Qurtoo's articles on tuning thinking budgets and trimming wasted reasoning tokens go further.

Example

Problem: Schedule six support staff across a week of morning and evening shifts.
Constraints:
- Each shift needs two people.
- Nobody works more than five shifts.
- Ana and Raj can't work evenings.
A correct answer is a table of shifts and names, followed by one line per constraint showing how it's met.
If the constraints can't all be met, say which ones conflict instead of producing a schedule.

When to use it

  • Multi-step math, planning, scheduling and code problems
  • Analysis where a wrong answer costs more than a slow one
  • Tasks where you can state what a correct answer looks like

When not to use it

  • Simple lookups, classification and rewriting, where the extra thinking only adds cost and delay
  • Tight latency or cost budgets
  • Workflows that depend on sampling settings the model doesn't accept