Reasoning Model.
A class of LLMs trained specifically to 'think' through problems step by step using extended chain-of-thought reasoning before producing a final answer, excelling at math, coding, and complex logic.
Reasoning models represent a significant capability leap over standard instruct models. Rather than directly generating an answer, they spend additional 'thinking tokens' working through problems step by step. OpenAI's o1 (released September 2024) was the first widely available reasoning model, followed by o3, and Anthropic's Claude 3.7 Sonnet with extended thinking.
These models are typically trained using reinforcement learning on problems with verifiable correct answers — mathematics, competitive programming, and formal logic — where the model receives a signal only when it reaches the right final answer. This forces it to develop useful intermediate reasoning rather than pattern-matching to surface-level features.
Reasoning models come with trade-offs: they are slower (more tokens = more time and cost), but dramatically more accurate on hard problems. On PhD-level science questions (GPQA Diamond), o3 scores above 85% — approaching human expert performance. The frontier is moving rapidly toward models that can dynamically allocate reasoning effort based on problem difficulty.
Related terms.
- Chain of ThoughtA prompting technique that encourages an LLM to reason step by step before giving a final answer, dramatically improving performance on complex tasks.
- RLHFReinforcement Learning from Human Feedback — a training technique that uses human preferences to teach AI models to be helpful, honest, and harmless.