Skip to slide
27 / 53

CH 04 · DeepSeek-R1, Learning to Reason Through Reinforcement Learning · 01/07

Read as one page

DeepSeek-R1, Learning to Reason Through Reinforcement Learning

DeepSeek-R1 rewards a model only for correct, verifiable answers. Its R1-Zero variant taught itself long reasoning and self-checking with no human examples.

7 slides · 3 min read

Contour map for DeepSeek-R1, Learning to Reason Through Reinforcement Learning

In this chapter

  1. 01The setup: reward correct answers, nothing else
  2. 02DeepSeek-R1-Zero: reasoning appears on its own
  3. 03The catch, and the fix: DeepSeek-R1
  4. 04Sharing the ability: distillation
  5. 05Why this paper mattered
  6. 06The one-sentence takeaway

Paper: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (2025)

So far, models learned to reason mostly by imitation: we showed them step-by-step examples (Chapter 1) or human-labeled good and bad steps (Chapter 3). DeepSeek-R1 asks a bolder question. What if we do not show the model how to reason at all, and instead just reward it for getting answers right, letting it discover how to reason on its own? The result was an open-source model that rivals the best closed reasoning models, and a genuinely surprising finding about how reasoning can emerge.

Finished this idea?

Your learning progress stays in this browser.