> ML_LITERATURE // SHAO-2024-DEEPSEEKMATH-GRPO-REASONING_v1.0
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, Daya Guo · arXiv preprint (2024)
algorithm2024industry-standardthirdPartyReproduced
Principal Contribution
Invented Group Relative Policy Optimization (GRPO), eliminating the critic/value network in PPO by computing group baseline averages over sampled responses.
Operational Relevance
Directly guides architectural decisions, alignment strategy, and serving infrastructure for task-text-generation, task-code-generation.
Assumptions
- Empirical distribution regularity holds and target domain adheres to pretraining linguistic/visual support
Limitations
- Resource scaling, inference memory requirements, and alignment robustness vary with model size and hardware topology
Connected Algorithms, Architectures & Tools
Related Algorithms:
Related Architectures:
Implementing Libraries: