Reproduce “Arrows of Math Reasoning Data Synthesis for Large Language Models: Diversity, Complexity and Correctness” 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
Sirui Chen, Changxin Tian, Binbin Hu, Kunlong Chen, Ziqi Liu, Zhiqiang Zhang, Jun Zhou
Ant Group researchers developed a program-assisted data synthesis framework, AMD, which generates a high-quality mathematical corpus with guaranteed diversity, complexity, and correctness. This synthetic data significantly enhances the intrinsic mathematical reasoning abilities of large language models across various benchmarks.
Key insights
- Combining knowledge system-derived and seed corpus-derived data is crucial for achieving comprehensive coverage and empirical relevance in synthetic mathematical datasets.
- A bilateral verification mechanism, comparing LLM outputs with program execution, effectively guarantees the correctness and consistency of synthesized problem-solution pairs.
- Systematic program mutation strategies enable the generation of a large volume of diverse and complex mathematical problems from simpler origins.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.