← AI Terminology
Reasoning Model
A reasoning model is an LLM specialised or trained to allocate extended internal chain-of-thought (and often RL) before answering, optimised for multi-step problems rather than instant chat replies.
Examples include OpenAI o-series, DeepSeek-R1, and similar 'thinking' modes from major labs.
Examples include OpenAI o-series, DeepSeek-R1, and similar 'thinking' modes from major labs.
Why It Matters in AI
Standard chat models are tuned for helpful fluent answers; reasoning models are tuned for hard math, code, and science where intermediate scratch work helps. They change product UX (hidden or visible traces), pricing, and eval strategy.
Key Points
| Aspect | Description |
|---|---|
| UX | Summarised answers with optional full reasoning traces |
| Evals | GPQA, MATH, Codeforces-like, live coding benches |
| Tradeoff | Higher latency/cost; huge gains on STEM benchmarks |
| Training | Often SFT on long CoT + RL with verifiable rewards |
| Behaviour | Longer internal monologue, self-checks, less impulsivity |
| Vs base LLM | Same backbone family possible; different post-training |
Simple Analogy
A colleague who pauses, sketches a proof on a whiteboard, then speaks — versus one who answers immediately from first instinct.
Common Usage Examples
- OpenAI o1/o3 API reasoning tokens
- DeepSeek-R1 open reasoning traces
- Gemini Flash/Thinking modes
- Route easy chat to fast model, hard tasks to reasoning model
Summary
In short: Reasoning models are LLMs post-trained to think in long chains before answering — trading latency for large gains on multi-step hard problems.