Reasoning models are trained to improve performance on tasks that involve multiple steps. In DeepSeek’s R1 paper, reinforcement learning encouraged patterns such as verification and self-reflection. Training a model this way differs from requesting steps through chain-of-thought prompting.
Some models let you choose how much work to spend on an answer. At its February 2025 launch, Claude 3.7 Sonnet offered ordinary responses and an extended-thinking mode. Extended thinking was a choice, rather than a requirement for every request.
Additional test-time compute can improve results, but a longer explanation is not a guarantee of correctness. Compare answer quality, cost and latency on your task. The AI reasoning overview links the prompting, inference and faithfulness research behind these distinctions.
See AI vocabulary in plain English for this term alongside the other language of models, data, agents, and infrastructure.
Sources
- DeepSeek-R1: Incentivizing Reasoning via Reinforcement Learning — Primary evidence of training that encourages verification and self-reflection on the tasks studied.
- Claude 3.7 Sonnet launch, February 2025 — Historical example of selectable ordinary responses and extended thinking.
Go deeper
From the conversation
-
How DeepSeek Changed the AI Race Overnight -
Why LLMs Are Plausibility Engines, Not Truth Engines | Dan Klein, Scaled Cognition -
Hallucinations Are a Data Architecture Problem | Sudhir Hasbe, Neo4j -
Agent Memory: The Last Battleground in the AI Stack | Richmond Alake, Oracle -
The Making of Gemini 2.0: DeepMind's Approach to AI Development and Deployment | Logan Kilpatrick -
The Critical Infrastructure Behind the AI Boom | Cisco CPO Jeetu Patel -
Beyond Transformers: How Liquid AI Is Rethinking LLM Architecture | Maxime Labonne -
Mastering Multi-Agent Systems | MongoDB’s Mikiko Chandrasekhar