ProRL Agent boosts SWE‑Bench
ProRL Agent, which decouples rollouts as RaaS, reportedly doubled SWE‑Bench scores for Qwen3‑8B from 9.6% to 18% — a measurable developer‑tool improvement. The same thread also surfaced an Open Physical AI Data Factory Blueprint for robotics and AV, signaling infrastructure focused on physical training data. (x.com)
The ProRL Agent paper on arXiv (ID 2603.18815) lists authors including Hao Zhang, Mingjie Liu and Jan Kautz and was posted in March 2026 as an open description of the rollout-as-a-service architecture. (arxiv.org) The system exposes rollouts via an HTTP service and coordinates an asynchronous three-stage “assembly line” of INIT, RUN and EVAL workers to overlap initialization, execution and scoring for high throughput. (marktechpost.com) To support shared HPC clusters the project uses Singularity for rootless sandboxing and reports micro-optimizations that cut shell-command latency from about 0.78s to 0.42s in their pipeline measurements. (marktechpost.com) NVIDIA has published the ProRL Agent server code under NVIDIA-NeMo/ProRL-Agent-Server on GitHub, describing a pluggable AgentHandler interface and OpenHands integration for high-concurrency rollout handling. (github.com) The authors validated ProRL Agent on SWE‑Bench Verified and report consistent gains across Qwen3 model scales (4B, 8B and 14B), with near-linear scaling behavior as compute resources increase. (arxiv.org) Separately, NVIDIA announced a Physical AI Data Factory Blueprint on March 16, 2026, framing it as an open reference architecture to automate generation, augmentation and evaluation of training data for robotics, vision agents and autonomous vehicles. (nvidianews.nvidia.com) NVIDIA named cloud partners including Microsoft Azure and Nebius and listed early industry adopters such as FieldAI, Hexagon Robotics, Linker Vision, Milestone Systems, Skild AI, Uber and Teradyne Robotics as users of the Data Factory blueprint. (investor.nvidia.com)
