NVIDIA launches Polar AI framework, boosting Codex benchmark scores by 594.74%
AI summary
NVIDIA's research team has open-sourced Polar, an agent reinforcement learning framework. It requires no rewriting of existing agent execution frameworks (such as Codex CLI, Claude Code, Qwen Code and Pi), instead plugging into GRPO training by placing an agent at the model API boundary. Experiments show that with the Qwen3.5-4B model, Polar raised Codex's pass@1 score on SWE-Bench Verified from 3.8% to 26.4% (a 594.74% increase). On efficiency, its prefix_merging technique cut training steps from 1,185 to 218, a roughly 5.39x speedup, while average GPU utilization rose from 20.4% to 87.7%.
Key Takeaways
- NVIDIA's research team has open-sourced Polar,an agent reinforcement learning framework.
- It requires no rewriting of existing agent execution frameworks (such as Codex CLI,Claude…
- Experiments show that with the Qwen3.
- 5-4B model,Polar raised Codex's pass@1 score on SWE-Bench Verified from 3.
- 74% increase).
Quick Facts
- Published
- Updated
- Source
- IT之家·人工智能
- Category
- Tutorials & Tips
- AI score
- 73
Polar 把 Codex 的 SWE-Bench 分数从 3.8% 拉到 26.4%,不是靠新模型而是靠训练框架,做代码 agent 的团队可以直接用,开源即拿即训。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: NVIDIA launches Polar AI framework, boosting Codex benchmark scores by 594.74%. May 28, 2026. https://new.nat6.net/en/news/cmpox6xm50a12slv4sw2e2viy/