Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning

The article presents Adaptive Data Scheduling (ADS), a dual-level framework designed to enhance reinforcement learning (RL) post-training for Large Language Models (LLMs) by replacing uniform data sampling with an adaptive approach based on semantic clusters and policy boundaries. Experimental results show that ADS improves average accuracy by 5.2% over Group Relative Policy Optimization (GRPO) across three LLMs and seven reasoning benchmarks, indicating its effectiveness as a versatile data scheduling strategy for practitioners in LLM reinforcement learning. The source code for ADS is publicly available on GitHub.