Reproduce “One Sample to Rule Them All: Extreme Data Efficiency in Multidiscipline Reasoning with Reinforcement Learning” 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
Yiyuan Li, Zhen Huang, Yanan Wu, Weixun Wang, Xuefeng Li, Yijia Luo, Wenbo Su, Bo Zheng and 1 more
Polymath learning demonstrates that training large language models with a single, carefully engineered sample via reinforcement learning can enhance multidisciplinary reasoning across diverse domains. This approach achieved performance comparable to or surpassing comprehensive learning with thousands of samples, promoting a shift towards "sample engineering" for extreme data efficiency in LLM development.
Key insights
- A single, carefully chosen mathematical reasoning sample can induce significant performance improvements not only in math but also across diverse, semantically distant domains like physics, biology, and literature.
- The breadth and coverage of foundational mathematical skills, particularly algebra and precalculus, within a single sample are key predictors of stronger multidisciplinary reasoning gains.
- Instruction-based synthesis can effectively engineer hybrid "meta-samples" that integrate multiple reasoning skills and multidisciplinary contexts, yielding superior cross-domain generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.