DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
by DeepSeek · free · Last verified 2026-03-17
This paper introduces GRPO (Group Relative Policy Optimization), a memory-efficient RL algorithm that replaces the critic model with group-sampled reward baselines, enabling scalable RLHF-style training. GRPO was subsequently adopted for DeepSeek-R1 and other reasoning-focused models.
https://arxiv.org/abs/2402.03300 ↗B
B—Above Average
Adoption: B+Quality: AFreshness: ACitations: B+Engagement: F
Specifications
- License
- Open Access
- Pricing
- free
- Capabilities
- policy-optimization, math-reasoning, rl-without-critic, scalable-rlhf
- Integrations
- Use Cases
- mathematical-reasoning, llm-fine-tuning, rl-training
- API Available
- No
- Tags
- reinforcement-learning, grpo, math-reasoning, deepseek, policy-optimization
- Added
- 2026-03-17
- Completeness
- 100%
Index Score
66.1Adoption
72
Quality
89
Freshness
88
Citations
78
Engagement
0
Put AI to work for your business
Deploy this paper alongside autonomous AaaS agents that handle tasks end-to-end — no babysitting required.