DeepSeek's V4 technical report indicates that the Group Relative Policy Optimization (GRPO) algorithm continues to be used for training domain-specific models. However, the overall post-training pipeline has evolved, with GRPO now being applied after supervised fine-tuning and before an on-policy distillation stage that consolidates capabilities into a unified model. This contrasts with the earlier V3.2 approach, which incorporated GRPO within a mixed reinforcement learning stage. AI
IMPACT Updates to GRPO's integration in DeepSeek's pipeline may influence how domain-specific models are trained and distilled into unified systems.
RANK_REASON The item discusses an evolution in an AI algorithm's application within a specific model's training pipeline, referencing technical reports and papers. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →