PEARL: Adaptive Prefill-Decode Execution with Elasticity for Agentic Reinforcement Learning

arXiv:2609.35158 · cs.AI · Submitted 2026-09-28 · Read on arXiv

cs.AI

Submitted: 2026-09-28

Updated: 2026-09-28

Code: https://github.com/NVIDIANeMo/RL

Terminology

• Decode — 1× in this paper · explained in 2 episodes
• Reinforcement Learning — 1× in this paper · explained in 4 episodes

Sources

• SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills
• Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
• DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
• ROSE: Rollout On Serving GPUs via Cooperative Elasticity for Agentic RL
• GLM-5: from Vibe Coding to Agentic Engineering
• Efficient Multi-round LLM Inference over Disaggregated Serving
• Not All Prefills Are Equal: PPD Disaggregation for Multi-turn LLM Serving
• DOPD: A Dynamic PD-Disaggregation Architecture for Maximizing Goodput in LLM Inference Serving
• InfiGUI-R1: Advancing Multimodal GUI Agents from Reactive Actors to Deliberative Reasoners
• GUI-R1 : A Generalist R1-Style Vision-Language Action Model For GUI Agents
• Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism
• Bidirectional Resource Scheduling for Disaggregated and Asynchronous RL Post-Training
• Kimi K2: Open Agentic Intelligence
• Kimi k1.5: Scaling Reinforcement Learning with LLMs
• INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning
• SeamlessFlow: A Trainer Agent Isolation RL Framework Achieving Bubble-Free Pipelines via Tag Scheduling
• Reinforcement Learning Optimization for Large-Scale Learning: An Efficient and User-Friendly Scaling Library
• RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning
• Qwen3 Technical Report
• TensorHub: Scalable and Elastic Weight Transfer for LLM RL Training

Related papers

• MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
• Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
• Agents' Last Exam
• The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
• MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
• Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems