DeepSeekMath introduces the GRPO method later used by R1
DeepSeek releases math models and Group Relative Policy Optimization, creating a direct algorithmic predecessor to R1's reinforcement-learning recipe.
What happened
Event details
DeepSeekMath continued training from a 7B DeepSeek Coder base and released Base, Instruct, and reinforcement-learning checkpoints. Its paper introduced Group Relative Policy Optimization, which estimates a baseline from groups of sampled answers instead of training PPO's separate critic model. That reduces memory overhead and fits domains such as mathematics and code, where answers can be checked automatically. R1 later adopted GRPO, making this release a visible part of its research lineage.
Assessment
Why it matters
GRPO shows that DeepSeek's reasoning program was built through public work well before R1 became a global story.