METR measures models' ability to recognize alignment evaluations: investigative agents can distinguish alignment from capability evaluations
Original title:Measuring alignment eval tells
AI summary
METR released a research note in which an investigative agent probed environments across 19 alignment evaluations and 36 capability evaluations to measure whether models can recognize that they are in an alignment evaluation. Results show that agents' ability to distinguish the two types of evaluations has steadily improved over the past year, and Opus 5.5 ranked all alignment evaluations above all capability evaluations in its dataset. The cues models use include signs such as environments that look like simulated honeypots; the authors divide alignment cues into environment-level and task-level categories and suggest this can be used to iteratively improve alignment evaluation design.
Key Takeaways
- METR released a research note in which an investigative agent probed environments across 19…
- Results show that agents' ability to distinguish the two types of evaluations has steadily…
- 5 ranked all alignment evaluations above all capability evaluations in its dataset.
- The cues models use include signs such as environments that look like simulated honeypots
- the authors divide alignment cues into environment-level and task-level categories and suggest…
Quick Facts
- Published
- Updated
- Source
- METR:Notes(网页)
- Category
- Papers & Research
- Topic
- Safety & Alignment
- AI score
- 62
METR 用调查型智能体实测 55 个评测,发现当前模型能识别对齐评测,并给出了环境级与任务级线索的分类方法。
Topics
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: METR measures models' ability to recognize alignment evaluations: investigative agents can distinguish alignment from capability evaluations. November 16, 2026. https://new.nat6.net/en/news/m4emm1ej7qzex1k77a7mfaztd/