Tools · MarkTechPost ·
NVIDIA Releases Polar, a Token-Faithful Rollout Framework for GRPO Training Across Codex, Claude Code, and Qwen Code
NVIDIA has released Polar, a token-faithful rollout framework for GRPO training that works across Codex, Claude Code, and Qwen Code harnesses. Using a proxy to capture token-level interactions, Polar improved SWE-Bench Verified pass@1 by up to 22.6 points on a Qwen3.5-4B model. The framework is available as a NeMo Gym