Post-training 0/0

Post-training

RL for reasoning

Reinforcement learning against rewards a program can check, such as passing tests or a correct final answer, used to train models that reason step by step, as in DeepSeek-R1.

Reading

  1. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning(opens in a new tab)

    Guo et al. 2025 Nature 60 min

    Reinforcement learning with rule-based rewards for correct final answers, applied to a base model with no supervised reasoning examples, produced long chains of thought that check and revise themselves (DeepSeek-R1-Zero). The released DeepSeek-R1 adds a small cold-start dataset and further training stages, and its reasoning was distilled into smaller open models.