Improving the Composer coding model with real-time reinforcement learning
Original title:Improving Composer through real-time RL
AI summary
The Cursor team applied real-time reinforcement learning to the Composer coding model, using reasoning tokens generated from real user interactions as training signals to address the mismatch between simulated environments and real-world use. The technique enables the team to deploy improved model checkpoints every five hours. Through A/B testing, the new version achieved improvements in key metrics: persistence of agent edits in the codebase increased by 2.28%, user dissatisfaction follow-ups decreased by 3.13%, and latency dropped by 10.3%. Real-time RL also brings new challenges such as reward hacking, but real user feedback helps identify and correct such issues.
Key Takeaways
- The Cursor team applied real-time reinforcement learning to the Composer coding model,using…
- The technique enables the team to deploy improved model checkpoints every five hours.
- Through A/B testing,the new version achieved improvements in key metrics: persistence of agent…
- 28%,user dissatisfaction follow-ups decreased by 3.
- 13%,and latency dropped by 10.
Quick Facts
- Published
- Updated
- Source
- Cursor Blog
- Category
- Opinion
- AI score
- 69
Cursor 把真实用户交互当训练信号,每五小时迭代一次 Composer,这不是论文是工程日志。做 coding agent 的团队该看看他们怎么处理 reward hacking 的两个真实案例,比任何 benchmark 论文都实在。
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: Improving the Composer coding model with real-time reinforcement learning. March 26, 2026. https://new.nat6.net/en/news/cmosbbgu40077slojavnrbfuj/