DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

TL;DR

DeepSeekMath combines a math-focused data pipeline with supervised fine-tuning and reinforcement learning, and introduces Group Relative Policy Optimization.

Why it matters

GRPO became a widely used reference for reasoning post-training because it removes the separate critic while retaining a policy-optimization formulation.

Key findings

1. 01

   Domain data construction and RL work together rather than as substitutes.

2. 02

   Group-relative baselines offer a practical alternative to critic-based PPO setups.

Scaling dimensions

Models, methods & benchmarks

Models

DeepSeekMath

Algorithms

GRPO

Topics