Reasoning Model
Also known as: Thinking model, o1-style model
An LLM trained to reason at length internally before answering. Slower and more expensive, but much better on hard problems.
An LLM trained (often via RL) to produce long internal chains of reasoning before its final answer. Examples: OpenAI o1/o3, DeepSeek-R1, Gemini Thinking. Trade higher latency and cost for much stronger performance on math, code, and multi-step problems.
In practice
The 2024-2025 paradigm shift. Senior interviews probe when reasoning models are worth the latency/cost and how their inference budget is controlled.
How it compares
Reasoning models bake CoT into RL training; CoT alone is just a prompting technique on any model.
Related topics
Questions that mention this term
- What distinguishes an AI agent from a plain LLM call?
- Describe the typical agent execution loop.
- Prompt-tuning vs LoRA: pick the architectural reason LoRA tends to win
- Diagnose the reasoning error in 'a bigger tokenizer vocabulary always cuts our token bill'
- What is the primary reason for running LLM judges at temperature=0 in an eval pipeline?
- A response comes back with finish_reason = 'length' which interpretation is correct?
Related terms
Chatbot Arena
Crowdsourced blind A/B comparisons of LLM responses, aggregated into an Elo leaderboard.
DPO (Direct Preference Optimization)
Skip the reward model and PPO: fine-tune directly on preferred-vs-rejected response pairs.
Eval Harness
A framework for running, scoring, and tracking LLM evals over time; handles golden sets, judges, and regressions.
Golden Set
Curated test cases with expected outputs; your regression suite for prompt and model changes.
GRPO (Group Relative Policy Optimization)
PPO without the critic: advantages are computed by ranking multiple sampled responses against each other.
Guardrails
Pre- and post-processing layers that block bad inputs/outputs and enforce policy on top of an LLM.