Reasoning model
A language model trained to think step by step before answering, for math, code and multi-step problems.
What it is
Reasoning models spend extra computation on a hidden or visible chain of thought before giving a final answer. That makes them slower and costlier per question, but much better at problems that need planning, checking and backtracking, such as math proofs, hard coding tasks and analysis.
How it works
A base LLM is trained with reinforcement learning on problems whose answers can be checked automatically (math results, passing tests). The model is rewarded for reaching correct answers, and learns to write long intermediate reasoning, try alternatives and verify its work. At run time, more thinking tokens usually means better answers.
What it is good for
- Math, science and logic problems
- Difficult coding and debugging
- Planning steps for agents
- Careful analysis where accuracy matters more than speed
Limits to know
- Slower and more expensive per answer
- Overkill for simple questions
- Long reasoning can still go wrong confidently
