Sandbagging (AI)

TagLast edit: 27 Mar 2025 18:20 UTC by Raemon

Sandbagging is when an AI system pretends to be less capable during training/evaluation.

Notes on countermeasures for exploration hacking (aka sandbagging)

ryan_greenblatt24 Mar 2025 18:39 UTC

52 points

6 comments8 min readLW link

The “no sandbagging on checkable tasks” hypothesis

Joe Carlsmith31 Jul 2023 23:06 UTC

56 points

14 comments9 min readLW link

Automated Researchers Can Subtly Sandbag

gasteigerjo, Akbir Khan, Sam Bowman, Vlad Mikulik, Ethan Perez and Fabien Roger

26 Mar 2025 19:13 UTC

44 points

0 comments4 min readLW link

(alignment.anthropic.com)

Won’t vs. Can’t: Sandbagging-like Behavior from Claude Models

Joe Benton and Zachary Witten

19 Feb 2025 20:47 UTC

15 points

1 comment1 min readLW link

(alignment.anthropic.com)

A Solution to Sandbagging and other Self-Provable Misalignment: Constitutional AI Detectives

Knight Lee14 Apr 2025 10:27 UTC

−3 points

2 comments4 min readLW link

An Introduction to AI Sandbagging

Teun van der Weij, Felix Hofstätter and Francis Rhys Ward

26 Apr 2024 13:40 UTC

46 points

13 comments8 min readLW link

How to mitigate sandbagging

Teun van der Weij23 Mar 2025 17:19 UTC

23 points

0 comments8 min readLW link

[Paper] AI Sandbagging: Language Models can Strategically Underperform on Evaluations

Teun van der Weij, Felix Hofstätter, Ollie J, Sam F. Brown and Francis Rhys Ward

13 Jun 2024 10:04 UTC

84 points

10 comments2 min readLW link

(arxiv.org)

Can SAE steering reveal sandbagging?

jordine, Hoang Khiem, Felix Hofstätter and Cleo Nardo

15 Apr 2025 12:33 UTC

35 points

3 comments4 min readLW link

No comments.