NVIDIA Releases Polar, a Token-Faithful Rollout Framework for GRPO Training Across Codex, Claude Code, and Qwen Code

NVIDIA released Polar, a proxy-based rollout framework for training language agents with reinforcement learning.

Polar captures prompt and response tokens, reconstructs trajectories, and preserves native harness behavior without rewrites.

The system is streaming-compatible and routes sessions through rollout servers and gateway nodes.

It includes support for tools like Codex CLI, Claude Code, Qwen Code, Gemini CLI, OpenCode, and Pi.

Built-in evaluators target session completion and SWE-Bench/SWE-Gym-style tasks, lowering the cost of GRPO-style training.