R1-Zero was the first to demonstrate that large language models can autonomously develop powerful reasoning capabilities entirely through reinforcement learning (RL), without requiring any supervised fine-tuning (SFT) data as a cold start.
The study uses the DeepSeek-V3-Base (671B parameters) foundation model, applying the innovative GRPO algorithm and a rule-based reward system for pure RL training. During training, the model exhibits "aha moments," autonomously learning complex reasoning behaviors such as reflection, verification, and generating long chains of thought.
On the AIME 2024 math competition, its accuracy surged from the base model's 15.6% to 71.0% (reaching 86.7% with majority voting), achieving performance on par with OpenAI o1-0912.
This result challenges the traditional paradigm that relies on large volumes of labeled data, demonstrating that reasoning capability may be an intrinsic property of models — one that can be elicited through appropriate reward signals alone. It opens a new path toward reducing data dependency and enabling more autonomous AI capability development.
Despite limitations such as poor readability and language mixing, the success of R1-Zero directly inspired the development of the full DeepSeek-R1 version, optimized through multi-stage training.