SPEED-RL: Faster Training of Reasoning Models via Online Curriculum Learning

TL;DR

SPEED-RL introduces an adaptive curriculum method that selects intermediate difficulty prompts during training, doubling or tripling training speed for reasoning models without sacrificing accuracy.

Contribution

It proposes a novel online curriculum learning approach that adaptively selects prompts of intermediate difficulty, improving training efficiency in reinforcement learning for language models.

Findings

Achieves 2x to 6x faster training times

Maintains model accuracy while improving efficiency

No manual tuning required for the curriculum method

Abstract

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.