Reproduce “Understanding Curriculum Learning in Large Language Models via Cross-Difficulty Optimization Dynamics” 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

Zhikai Ding, Ziyi Ye

Zhikai Ding and Ziyi Ye from Fudan University introduce "Relative Transfer," an optimization-based measure quantifying how training on one difficulty level affects others in large language models. This work demonstrates that an adaptive sampling strategy, Transfer-aware Dynamic Curriculum Sampling (TDCS), which dynamically adjusts data distribution based on this transfer, consistently improves performance across various reasoning benchmarks and model scales, achieving up to 3.8% higher accuracy.

Key insights

  • Cross-difficulty knowledge transfer, quantified by "Relative Transfer," mechanistically explains the varying effectiveness of fixed curriculum schedules across tasks.
  • Tasks with strong positive transfer from harder to easier examples benefit from pure easy-to-hard curricula (e.g., Sudoku), while tasks with weaker transfer require dynamic data replay (e.g., iGSM, KodCode).
  • Dynamically adjusting data sampling based on continuous estimation of Relative Transfer enables a more robust and effective curriculum strategy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.