NVIDIA has unveiled ProRL AGENT, a decoupled Rollout-as-a-Service infrastructure for training multi-turn AI agents. The ...

NVIDIA has unveiled ProRL AGENT, a decoupled Rollout-as-a-Service infrastructure for training multi-turn AI agents. The system separates I/O-heavy environment interactions from GPU-intensive policy training, running them as independent services with a three-stage async pipeline. On SWE-bench Verified, Qwen3-8B improved from 9.6% to 18.0%, Qwen3-14B jumped from 15.4% to 23.6%. 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 #AIInfrastructure #NVIDIA

Read Original

Related