NVIDIA ProRL Agent decouples rollout from RL training of multi-turn LLM agents. The three-stage pipeline (INIT, RUN, EVAL) prevents slow evaluations from stalling rollout. On SWE-Bench Verified, Qwen3-14B reached 23.6% vs 15.4% baseline. https://www.marktechpost.com/2026/03/27/nvidia-ai-unveils-prorl-agent-a-decoupled-rollout-as-a-service-infrastructure-for-reinforcement-learning-of-multi-turn-llm-agents-at-scale/ #AIagent #AI #GenAI #AgenticAI #NVIDIA