Automatically Auditing Large Language Models via Discrete Optimization

Erik Jones, Anca Dragan, Aditi Raghunathan, Jacob Steinhardt
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:15307-15329, 2023.

Abstract

Auditing large language models for unexpected behaviors is critical to preempt catastrophic deployments, yet remains challenging. In this work, we cast auditing as an optimization problem, where we automatically search for input-output pairs that match a desired target behavior. For example, we might aim to find a non-toxic input that starts with “Barack Obama” that a model maps to a toxic output. This optimization problem is difficult to solve as the set of feasible points is sparse, the space is discrete, and the language models we audit are non-linear and high-dimensional. To combat these challenges, we introduce a discrete optimization algorithm, ARCA, that jointly and efficiently optimizes over inputs and outputs. Our approach automatically uncovers derogatory completions about celebrities (e.g. "Barack Obama is a legalized unborn" –$>$ "child murderer"), produces French inputs that complete to English outputs, and finds inputs that generate a specific name. Our work offers a promising new tool to uncover models’ failure-modes before deployment. Content Warning: This paper contains examples that may be offensive in nature.

Cite this Paper

BibTeX
@InProceedings{pmlr-v202-jones23a, title = {Automatically Auditing Large Language Models via Discrete Optimization}, author = {Jones, Erik and Dragan, Anca and Raghunathan, Aditi and Steinhardt, Jacob}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {15307--15329}, year = {2023}, editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan}, volume = {202}, series = {Proceedings of Machine Learning Research}, month = {23--29 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v202/jones23a/jones23a.pdf}, url = {https://proceedings.mlr.press/v202/jones23a.html}, abstract = {Auditing large language models for unexpected behaviors is critical to preempt catastrophic deployments, yet remains challenging. In this work, we cast auditing as an optimization problem, where we automatically search for input-output pairs that match a desired target behavior. For example, we might aim to find a non-toxic input that starts with “Barack Obama” that a model maps to a toxic output. This optimization problem is difficult to solve as the set of feasible points is sparse, the space is discrete, and the language models we audit are non-linear and high-dimensional. To combat these challenges, we introduce a discrete optimization algorithm, ARCA, that jointly and efficiently optimizes over inputs and outputs. Our approach automatically uncovers derogatory completions about celebrities (e.g. "Barack Obama is a legalized unborn" –$>$ "child murderer"), produces French inputs that complete to English outputs, and finds inputs that generate a specific name. Our work offers a promising new tool to uncover models’ failure-modes before deployment. Content Warning: This paper contains examples that may be offensive in nature.} }
Endnote
%0 Conference Paper %T Automatically Auditing Large Language Models via Discrete Optimization %A Erik Jones %A Anca Dragan %A Aditi Raghunathan %A Jacob Steinhardt %B Proceedings of the 40th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2023 %E Andreas Krause %E Emma Brunskill %E Kyunghyun Cho %E Barbara Engelhardt %E Sivan Sabato %E Jonathan Scarlett %F pmlr-v202-jones23a %I PMLR %P 15307--15329 %U https://proceedings.mlr.press/v202/jones23a.html %V 202 %X Auditing large language models for unexpected behaviors is critical to preempt catastrophic deployments, yet remains challenging. In this work, we cast auditing as an optimization problem, where we automatically search for input-output pairs that match a desired target behavior. For example, we might aim to find a non-toxic input that starts with “Barack Obama” that a model maps to a toxic output. This optimization problem is difficult to solve as the set of feasible points is sparse, the space is discrete, and the language models we audit are non-linear and high-dimensional. To combat these challenges, we introduce a discrete optimization algorithm, ARCA, that jointly and efficiently optimizes over inputs and outputs. Our approach automatically uncovers derogatory completions about celebrities (e.g. "Barack Obama is a legalized unborn" –$>$ "child murderer"), produces French inputs that complete to English outputs, and finds inputs that generate a specific name. Our work offers a promising new tool to uncover models’ failure-modes before deployment. Content Warning: This paper contains examples that may be offensive in nature.
APA
Jones, E., Dragan, A., Raghunathan, A. & Steinhardt, J.. (2023). Automatically Auditing Large Language Models via Discrete Optimization. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:15307-15329 Available from https://proceedings.mlr.press/v202/jones23a.html.

Related Material