New Framework Optimizes LLM Curriculum Learning

This research addresses a key challenge in large language model (LLM) training: understanding why curriculum learning, which organizes data from easy to hard, has inconsistent effectiveness across different reasoning tasks. The paper proposes that the varying success is due to the transfer relationship between different difficulty levels during optimization. To quantify this, the authors introduce "Relative Transfer," a new measure that characterizes cross-difficulty knowledge transfer. Based on this insight, they developed Transfer-aware Dynamic Curriculum Sampling (TDCS). TDCS is an adaptive framework that dynamically adjusts the sampling distribution of training data throughout the training process, guided by the estimated transfer relationships. Extensive experiments across multiple reasoning benchmarks, model scales, and training paradigms demonstrate that TDCS consistently outperforms existing curriculum scheduling strategies. This work provides a unified, optimization-based explanation for curriculum learning's effectiveness, highlighting the importance of understanding knowledge transfer across task difficulties.