NVIDIA releases open-source Polar to boost RL training for coding agents

NVIDIA researchers released Polar, an open-source framework that helps existing coding agent tools like Codex, Claude Code, and Qwen Code connect to GRPO training without changing how they work. GRPO is a method that improves model behavior through rewards. Polar focuses on the interface between agents and models instead of rewriting the whole execution system. In tests using Qwen3.5-4B, Codex's pass@1 score jumped from 3.8% to 26.4% on SWE-Bench Verified, a nearly 6x improvement. Polar also makes training faster by reducing update count by 80% and cutting wall-clock time by 5.39x, while boosting GPU usage from 20.4% to 87.7%. It uses two main parts: rollout server for task management and gateway node for session execution.