PEARL: Adaptive Prefill-Decode Execution with Elasticity for Agentic Reinforcement Learning
PEARL adapts prefill-decode execution and elastic GPUs, raising agentic RL throughput by up to 2.79x.
PEARL is an asynchronous agentic reinforcement-learning system that combines external GPU elasticity, temporary reuse of idle training GPUs, and adaptive prefill-decode execution. Runtime profiles predict rollout completion, including environment-driven drops in decode concurrency, and cost-aware policies limit low-benefit reconfiguration. It achieves 2.17-2.79x the throughput of fixed-resource ROLL and improves over RLBoost+ by up to about 26.9% for Qwen3-8B and 36.3% for Qwen3-30B-A3B.
- Coordinates elastic GPUs with adaptive prefill-decode execution
- Temporarily reuses idle training GPUs between updates
- Delivers 2.17-2.79x the throughput of fixed-resource ROLL
- Improves on RLBoost+ by up to 36.3% for Qwen3-30B-A3B
- AI models
- Qwen3-8BQwen3-30B-A3B
From the original article
1 min read at arxiv.orgMulti-turn rollout dominates the cost of agentic reinforcement learning (RL). Asynchronous execution and elastic GPU resources can accelerate this stage, but adding rollout replicas yields diminishing returns while training GPUs remain idle between updates. We observe that effective resource use also depends on the prefill--decode (PD) configuration.
By Jiaan Zhu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.