Reproduce “What Makes a Good Curriculum? Disentangling the Effects of Data Ordering on LLM 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
Yaning Jia, Chunhui Zhang, Xingjian Diao, Xiangchi Yuan, Zhongyu Ouyang, Chiyu Ma, Soroush Vosoughi
Dartmouth College researchers systematically investigated the impact of data ordering in curriculum learning on mathematical reasoning in Large Language Models using an offline evaluation framework. They found that optimal curriculum strategies, varying by ordering direction and difficulty metric, depend on both the model's capabilities and the task's complexity, leading to performance gains of up to 3-6% on easier mathematical benchmarks.
Key insights
- No universal curriculum strategy exists; optimal data ordering depends jointly on the LLM's reasoning capacity, the complexity of the task, and the chosen difficulty metric.
- Curricula reshape optimization trajectories: internal-state metrics influence convergence speed and stability, while task-aligned metrics impact final convergence point and generalization.
- Curriculum learning modulates internal model states, with Forward CL tending to yield more cautious, uncertainty-aware models and Reverse CL often preserving decisiveness with higher confidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.