EasyDeepLearn

GRPO — Group Relative Policy Optimization (DeepSeekMath, R1).

hard

Answer

  • PPO variant that skips the value network.
  • For each prompt, sample G responses, compute rewards, normalize as advantages: Ai  =  (ri    meanr)  /  stdrA_{i}\; = \;(r_{i}\; - \;\mathrm{mean}_{r})\; / \;\mathrm{std}_{r}.
  • Uses group-normalized rewards as advantage estimates → no critic needed.
  • Preserves PPO's clipped surrogate loss.
  • Enables efficient RL fine-tuning of reasoning models (DeepSeek-R1).
  • Modern efficiency win for LLM RL.
Check yourself — multiple choice
  • Same as PPO
  • Sample G responses per prompt, normalize rewards as advantages (Ai  =  (ri    mean)/std)(A_{i}\; = \;(r_{i}\; - \;\mathrm{mean}) / \mathrm{std}); no critic needed; foundation of DeepSeek-R1 reasoning RL
  • Random
  • Not real

GRPO: group-normalized rewards as advantages; no critic; DeepSeek-R1.

#llm#alignment#policy-methods

Practise Reinforcement Learning

214 interview questions in this topic.

Related questions