Reproduce “ScaleDiff: Scaling Difficult Problems for Advanced Mathematical Reasoning” 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
Qizhi Pei, Zhuoshi Pan, Honglin Lin, Xin Gao, Yu Li, Zinan Tang, Conghui He, Rui Yan and 1 more
Researchers from OpenDataLab, Shanghai Artificial Intelligence Laboratory developed ScaleDiff, an efficient and scalable pipeline for generating difficult mathematical problems. Fine-tuning a 7B-scale model on this augmented data, ScaleDiff-7B achieved 65.9% average accuracy on complex benchmarks, an 11.3% relative gain, demonstrating that scaling difficult training data improves complex mathematical reasoning.
Key insights
- Training Large Reasoning Models (LRMs) on truly difficult problems is crucial for developing sophisticated, multi-step reasoning capabilities.
- Efficiently identifying genuinely difficult problems through adaptive thinking models, rather than costly methods, is key to curating high-value training data.
- Scaling the quantity of appropriately challenging training data leads to consistent performance improvements in LRMs on complex tasks, demonstrating a "scaling phenomenon" for difficulty.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.