Selector-Guided Curriculum Boosts One-Shot RLVR for LLMs

A new approach called Selector-Guided Autonomous Curriculum (SGAC) improves one-shot Reinforcement Learning from Verifiable Rewards (RLVR) for Large Language Models (LLMs). Current state-of-the-art methods use heuristics based on historical reward variance to select training instances, but this is misleading as a measure of transferability. SGAC employs a learnable selector model that considers a multi-dimensional feature space including success probability, reward variance, output disagreement (entropy), and semantic difficulty. Empirical evaluation shows output disagreement is the strongest predictor of reasoning gains, outperforming reward variance. The method was tested on pools of candidate problems.