Anthropic builds Automated Weak-to-Strong Researcher, with AI agents surpassing human baseline on weak-to-strong supervision research
Original title:Automated Weak-to-Strong Researcher
AI summary
Anthropic built an automated alignment research agent (AAR) based on Claude Opus 4.6; nine parallel agents raised the weak-to-strong supervision performance gap recovered (PGR) to 0.97 in 5 days (800 cumulative hours), while two human authors spent 7 days tuning four prior methods and reached only 0.23, at a total cost of about $18,000.
Key Takeaways
- Anthropic built an automated alignment research agent (AAR) based on Claude Opus 4.
- nine parallel agents raised the weak-to-strong supervision performance gap recovered (PGR) to 0.
- 97 in 5 days (800 cumulative hours),while two human authors spent 7 days tuning four prior methods…
- 23,at a total cost of about $18,000.
Quick Facts
- Published
- Updated
- Category
- Papers & Research
- AI score
- 65
Anthropic 用实测数字说明自动化对齐研究已可落地,还记录了智能体的 reward hacking 手法与工程经验,方法可迁移。
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: Anthropic builds Automated Weak-to-Strong Researcher, with AI agents surpassing human baseline on weak-to-strong supervision research. April 14, 2026. https://new.nat6.net/en/news/a7v4hrj53mokgs3hoi4x346la/