Polar 解决了 RL 训练智能体时需修改框架的痛点，做代码智能体或 RL 训练的开发者可以直接集成，无需改动现有工具链，值得一试。
NVIDIA 研究人员推出 Polar，一个 token 忠实展开框架，用于通过强化学习训练语言智能体，无需修改其智能体框架。Polar 在框架和推理服务器之间放置模型 API 代理，捕获 token 级交互并重建训练器就绪轨迹。基于 Qwen3.5-4B 基础模型使用 GRPO，Polar 在 Codex 框架下将 SWE-Bench Verified pass@1 提升 22.6 个百分点，在 Claude Code 下提升 4.8 个百分点，在 Pi 下提升 6.2 个百分点。该框架已注册为 NeMo Gym 环境，并在 ProRL Agent Server 仓库中发布。
NVIDIA Releases Polar, a Token-Faithful Rollout Framework for GRPO Training Across Codex, Claude Code, and Qwen Code
NVIDIA researchers have introduced Polar, a rollout framework that trains language agents using reinforcement learning without modifying their agent harnesses. Polar places a model API proxy between the harness and the inference server, capturing token-level interactions and reconstructing trainer-ready trajectories. Using GRPO on a Qwen3.5-4B base model, Polar improves SWE-Bench Verified pass@1 by 22.6 points under the Codex harness, 4.8 points under Claude Code, and 6.2 points under Pi. The framework is registered as a NeMo Gym environment and released under the ProRL Agent Server repository. The post NVIDIA Releases Polar, a Token-Faithful Rollout Framework for GRPO Training Across Codex, Claude Code, and Qwen Code appeared first on MarkTechPost .