SPEED-RL: Faster Training of Reasoning Models via Online Curriculum Learning
This addresses the problem of slow training times for AI researchers and practitioners developing reasoning models, offering a significant speedup with minimal overhead.
The paper tackles the computational expense of training large language models with reinforcement learning for reasoning by introducing SPEED, an adaptive online curriculum that selects intermediate-difficulty prompts, resulting in 2x to 6x faster training without accuracy loss.
Training large language models with reinforcement learning (RL) against verifiable rewards significantly enhances their reasoning abilities, yet remains computationally expensive due to inefficient uniform prompt sampling. We introduce Selective Prompting with Efficient Estimation of Difficulty (SPEED), an adaptive online RL curriculum that selectively chooses training examples of intermediate difficulty to maximize learning efficiency. Theoretically, we establish that intermediate-difficulty prompts improve the gradient estimator's signal-to-noise ratio, accelerating convergence. Empirically, our efficient implementation leads to 2x to 6x faster training without degrading accuracy, requires no manual tuning, and integrates seamlessly into standard RL algorithms.