Reproduce “Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu and 8 more
The Qwen Team at Alibaba Group developed the Qwen2.5-Math series, specialized large language models designed for mathematical reasoning. These models achieve state-of-the-art performance among open-source options, with Qwen2.5-Math-72B-Instruct scoring 85.9% on the MATH benchmark, by employing a novel self-improvement pipeline that iteratively refines the models through synthetic data generation and reward-guided reinforcement learning.
Key insights
- A self-improvement pipeline, leveraging synthetic data generation, an iterative reward model, and reinforcement learning, effectively reduces reliance on human-annotated data while continuously enhancing mathematical reasoning performance.
- Tool-Integrated Reasoning (TIR) consistently outperforms Chain-of-Thought (CoT) across benchmarks, highlighting the value of integrating external computational tools for complex mathematical problems.
- The Qwen2.5-Math series achieves strong performance-to-size ratios, with smaller models like Qwen2.5-Math-7B-Instruct reaching 83.6% accuracy on the MATH benchmark, making advanced mathematical reasoning more accessible.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.