GRPO - Group Relative Policy Optimization: How DeepSeek Trains Reasoning Models

Explore Group Relative Policy Optimization (GRPO), the reinforcement learning technique used by DeepSeek to train its advanced reasoning model, in this 22-minute educational video from Serrano.Academy. Discover how GRPO differs from self-supervised learning by using reinforcement learning for model self-improvement. Compare DeepSeek's reasoning capabilities with ChatGPT through practical examples, understand the GRPO scoring mechanism, and learn about key concepts including context-based answering, quality advantage, response probability calculations, and response clipping. The video breaks down complex technical concepts into accessible explanations, making it part of a broader series on reinforcement learning for large language models. Perfect for those interested in the technical foundations behind cutting-edge AI reasoning systems.