LESSWRONG
LW

Sandbagging (AI)

Edited by Raemon last updated 27th Mar 2025

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

Posts tagged Sandbagging (AI)

77Misalignment and Strategic Underperformance: An Analysis of Sandbagging and Exploration Hacking

Buck, Julian Stastny

4mo

3

54Notes on countermeasures for exploration hacking (aka sandbagging)

ryan_greenblatt

5mo

6

77White Box Control at UK AISI - Update on Sandbagging Investigations

Joseph Bloom, Jordan Taylor, Connor Kissane, Sid Black, merizian, alexdzm, jacoba, Ben Millwood, Alan Cooney

2mo

10

61The “no sandbagging on checkable tasks” hypothesis

2y

14

44Automated Researchers Can Subtly Sandbag

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

5mo

0

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

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

1y

10

50Do LLMs know what they're capable of? Why this matters for AI safety, and initial findings

Casey Barkan, Sid Black, Oliver Sourbut

2mo

4

48An Introduction to AI Sandbagging

Teun van der Weij, Felix Hofstätter, Francis Rhys Ward

1y

13

35Can SAE steering reveal sandbagging?

jordine, Hoang Khiem, Felix Hofstätter, Cleo Nardo

5mo

3

30How to mitigate sandbagging

Teun van der Weij

5mo

0

16Exploration hacking: can reasoning models subvert RL?

Damon Falck, Joschka Braun, Eyon Jang

1mo

4

15Won't vs. Can't: Sandbagging-like Behavior from Claude Models

Joe Benton, Zachary Witten

6mo

1

15Adding noise to a sandbagging model can reveal its true capabilities

2mo

1

-3A Solution to Sandbagging and other Self-Provable Misalignment: Constitutional AI Detectives

5mo

2

Add Posts