Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning
The Adaptive Data Scheduling (ADS) framework improves Large Language Model reinforcement learning by replacing uniform data sampling with a dual-level adaptive system. ADS first organizes data into semantic clusters and then adaptively samples between these clusters based on their success rates. It also selects specific 'policy-boundary' samples within those clusters to guide the policy toward its capability limits, leading to better accuracy across different reasoning tasks.