AI BriefingIntroducing Qwen2-Math
Key point
Qwen has released Qwen2-Math, a series of math-specialized LLMs that maximize mathematical problem-solving ability.
Details
Qwen2-Math and Qwen2-Math-Instruct (1.5B, 7B, 72B), new math-specialized LLMs in the Qwen series, have been released. These models show results that surpass the math performance of existing open-source models as well as major closed-source models such as GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Pro, and Llama-3.1-405B.
The Qwen2-Math base model is pretrained on Qwen2 using a math-specific corpus (web text, books, code, exam questions, and synthetic data). It has demonstrated excellent performance on major English benchmarks such as GSM8K, Math, and MMLU-STEM, as well as Chinese benchmarks such as CMATH and GaoKao.
The Qwen2-Math-Instruct model further improves performance by leveraging a math-specialized Reward Model.
- Applies reinforcement learning through Rejection Sampling and GRPO (Group Relative Policy Optimization)
- Delivers excellent performance even on high-difficulty math competition problems such as AIME 2024 and AMC 2023
- Achieves top-tier performance compared to models of the same scale
Currently, this model mainly supports English, and a bilingual math model supporting both English and Chinese is planned for future release.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.