Alignment & Preference Tuning · Fast-moving · Advanced
Group Relative Policy Optimization
Also known as: GRPO, Group RLHF
A reinforcement learning algorithm that optimizes policy outputs against a group of sampled completions without a critic model.
What Group Relative Policy Optimization is
GRPO reduces memory usage and training complexity during RLHF by computing relative advantages across a batch of completions.
How it works
Samples multiple responses per prompt, scores them with verifiable reward functions, and updates policy weights proportionally.
Why it matters
Pioneered in models like DeepSeek-Math and DeepSeek-R1 to scale mathematical reasoning capabilities efficiently.
Common uses
- →Reasoning model reinforcement learning
- →Math and code RLHF
- →Critic-free policy optimization
More in this collection
Browse all AI Concepts