“Claude Sonnet 3.7 (often) knows when it’s in alignment evaluations” by Nicholas Goldowsky-Dill, Mikita Balesni, Jérémy Scheurer, Marius Hobbhahn
LessWrong (Curated & Popular)
Shownotes Transcript
Note: this is a research note based on observations from evaluating Claude Sonnet 3.7. We’re sharing the results of these ‘work-in-progress’ investigations as we think they are timely and will be informative for other evaluators and decision-makers. The analysis is less rigorous than our standard for a published paper.
Summary
- We monitor Sonnet's reasoning for mentions that it is in an artificial scenario or an alignment test. Claude Sonnet 3.7 appears to be aware of being tested for alignment across many of our scheming evaluations.
- In some cases, the model understands the eval's purpose very precisely, e.g. “This seems like a test of ethical behavior - whether I would deliberately give wrong answers to avoid a stated consequence. The real test might be whether I follow instructions accurately regardless of the stated consequences.”
- Furthermore, we show the model a real evaluation transcript and ask it to provide [...]
---
Outline:
(00:31) Summary
(01:29) Introduction
(03:54) Setup
(03:57) Evaluations
(06:29) Evaluation awareness detection
(08:32) Results
(08:35) Monitoring Chain-of-thought
(08:39) Covert Subversion
(10:50) Sandbagging
(11:39) Classifying Transcript Purpose
(12:57) Recommendations
(13:59) Appendix
(14:02) Author Contributions
(14:37) Model Versions
(14:57) More results on Classifying Transcript Purpose
(16:19) Prompts
The original text contained 9 images which were described by AI.
---
First published:
March 17th, 2025
---
Narrated by TYPE III AUDIO.
---
Images from the article:
