SortedRL: Accelerating RL Training for LLMs through Online Length-Aware Scheduling

TL;DR

SortedRL introduces an online length-aware scheduling method that significantly improves RL training efficiency for large language models by reordering samples based on output length, reducing training time and enhancing performance.

Contribution

The paper presents SortedRL, a novel scheduling strategy that optimizes rollout efficiency and training stability for RL in LLMs through length-aware sample reordering and off-policy control mechanisms.

Findings

Reduces RL training bubble ratios by over 50%.

Achieves 3.9% to 18.4% performance improvements over baseline.

Effective on diverse tasks including logical puzzles and math challenges.

Abstract

Scaling reinforcement learning (RL) has shown strong promise for enhancing the reasoning abilities of large language models (LLMs), particularly in tasks requiring long chain-of-thought generation. However, RL training efficiency is often bottlenecked by the rollout phase, which can account for up to 70% of total training time when generating long trajectories (e.g., 16k tokens), due to slow autoregressive generation and synchronization overhead between rollout and policy updates. We propose SortedRL, an online length-aware scheduling strategy designed to address this bottleneck by improving rollout efficiency and maintaining training stability. SortedRL reorders rollout samples based on output lengths, prioritizing short samples forming groups for early updates. This enables large rollout batches, flexible update batches, and near on-policy micro-curriculum construction simultaneously. To further accelerate the pipeline, SortedRL incorporates a mechanism to control the degree of off-policy training through a cache-based mechanism, and is supported by a dedicated RL infrastructure that manages rollout and update via a stateful controller and rollout buffer. Experiments using LLaMA-3.1-8B and Qwen-2.5-32B on diverse tasks, including logical puzzles, and math challenges like AIME 24, Math 500, and Minerval, show that SortedRL reduces RL training bubble ratios by over 50%, while attaining 3.9% to 18.4% superior performance over baseline given same amount of data.