NVIDIA Releases AI Framework Polar, Raising Codex Benchmark Score by 594.74%

英伟达推出 AI 框架 Polar，让 Codex 跑分暴涨 594.74%

NVIDIA’s research team open-sourced Polar, an agent reinforcement learning framework that connects GRPO training at the model API boundary without rewriting Codex CLI, Claude Code, Qwen Code, or Pi; on Qwen3.5-4B, Polar raised Codex pass@1 on SWE-Bench Verified from 3.8% to 26.4%, while prefix_merging cut training steps from 1,185 to 218.

Why it matters: HKR-H/K/R all pass: NVIDIA open-sourced Polar with a concrete GRPO mechanism and SWE-Bench Verified numbers. This is a strong research/open-source item, not a major model or product release, so it stays in the 78–84 band.