New KPop method pushes Ring-2.6-1T past 76 on SWE-bench Verified
Original title:From IcePop to KPop — our team keeps pushing on RL training stability for large MoE models. 👇 KPop...
AI summary
The team introduced KPop to stabilize agentic reinforcement learning training for large-scale MoE models. It replaces the fixed-ratio masking used in the earlier IcePop method with an adaptive masking mechanism based on binary KL divergence, dynamically adjusting according to the degree of train-inference mismatch during training. This improvement enabled the Ring-2.6-1T model to score above 76 on SWE-bench Verified through pure RL training alone, without modifying infrastructure or routing replay.
Key Takeaways
- The team introduced KPop to stabilize agentic reinforcement learning training for large-scale MoE…
- It replaces the fixed-ratio masking used in the earlier IcePop method with an adaptive masking…
- This improvement enabled the Ring-2.
- 6-1T model to score above 76 on SWE-bench Verified through pure RL training alone,without…
Quick Facts
- Published
- Updated
- Source
- X:蚂蚁百灵 (@AntLingAGI)
- Category
- Papers & Research
- AI score
- 68
蚂蚁团队把 IcePop 升级成 KPop,从固定掩码变成自适应 KL 区域,思路很巧。Ring-2.6-1T 纯 RL 直接冲到 SWE-bench 76+,做 agentic RL 训练的同学值得翻一下博客。
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: New KPop method pushes Ring-2.6-1T past 76 on SWE-bench Verified. May 26, 2026. https://new.nat6.net/en/news/cmpmsjd610rghsl01l5zfielj/