# What is Group Relative Policy Optimization? Definition & Technical Mechanics

> **Source:** [SPIDITS AI AI Glossary](https://spidits.com/ai-glossary)  
> **Category:** Post-Training & Optimization  
> **Canonical Citation:** [https://spidits.com/ai-glossary/group-relative-policy-optimization](https://spidits.com/ai-glossary/group-relative-policy-optimization)

## Definition
Group Relative Policy Optimization (GRPO) is a memory-efficient reinforcement learning algorithm for LLM post-training that estimates baseline advantages by normalizing rewards across a group of sampled candidate outputs for a prompt, eliminating the need for a separate value critic model.

## Architectural Mechanics & Deep Dive
Group Relative Policy Optimization (GRPO) is an alignment and reinforcement learning algorithm popularized by DeepSeek (used to train DeepSeekMath and DeepSeek-R1). Traditional Proximal Policy Optimization (PPO) requires maintaining a dedicated Critic (value) network to estimate state-value baselines, consuming vast amounts of GPU VRAM. GRPO eliminates the Critic model entirely. For every prompt, GRPO samples a group of candidate outputs, calculates rewards for each, and computes the relative z-score baseline. This relative reward optimizes the policy directly with significantly reduced compute.

## Industry Applications & Real-World Use Cases
Reasoning model post-training, RLHF alignment, and memory-efficient policy optimization

## Frequently Asked Questions

### What is Group Relative Policy Optimization (GRPO)?
GRPO is a reinforcement learning algorithm created by DeepSeek that evaluates rewards relative to the average of a group of outputs, eliminating the critic model required by PPO.

### Why is GRPO more memory-efficient than PPO?
Standard PPO requires hosting a separate Critic model of equal size to the Policy model; GRPO eliminates the Critic entirely by deriving baselines directly from group relative scores.

## Related Terminology
- [dpo](https://spidits.com/ai-glossary/dpo)
- [rlhf](https://spidits.com/ai-glossary/rlhf)
- [grpo](https://spidits.com/ai-glossary/grpo)
- [deepseek](https://spidits.com/ai-glossary/deepseek)
- [process-reward-model](https://spidits.com/ai-glossary/process-reward-model)

---
*From the SPIDITS AI Knowledge Base (336+ terms indexed). Explore full technical definitions: [https://spidits.com/ai-glossary](https://spidits.com/ai-glossary)*