SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives

Vincent Siu, Nicholas Crispino, David Park, Nathan W. Henry, Zhun Wang, Yang Liu, Dawn Song, Chenguang Wang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:113832-113884, 2026.

Abstract

We introduce STEERINGSAFETY, a benchmark for evaluating representation steering methods across nine safety perspectives spanning 18 datasets. While prior work highlights the general capabilities of representation steering, we focus on safety perspectives including refusal, bias, hallucination, social behaviors, reasoning, epistemic integrity, and normative judgment. STEERINGSAFETY provides modularized building blocks for state-of-the-art steering methods, enabling unified implementation of DIM, ACE, CAA, PCA, and LAT with recent enhancements such as conditional steering. Results on Gemma-2-2B, Llama-3.1-8B, and Qwen-2.5-7B show that strong steering performance depends on the pairing of method, model, and specific perspective. For instance, DIM is consistently effective, yet all methods exhibit substantial entanglement, where improving effectiveness on one safety perspective often significantly changes performance on others. Social behaviors are most vulnerable (degradation up to 76%), refusal steering (jailbreaking) frequently compromises normative judgment such as commonsense morality (up to 26%), and hallucination steering shifts political views unpredictably across models, ranging from a 25% shift to the right to a 28% shift to the left. These findings show the need to understand steering methods through multiple safety angles rather than a single target behavior.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-siu26a, title = {{S}teering{S}afety: Benchmarking Representation Steering in {LLM}s Across Safety Perspectives}, author = {Siu, Vincent and Crispino, Nicholas and Park, David and Henry, Nathan W. and Wang, Zhun and Liu, Yang and Song, Dawn and Wang, Chenguang}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {113832--113884}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/siu26a/siu26a.pdf}, url = {https://proceedings.mlr.press/v306/siu26a.html}, abstract = {We introduce STEERINGSAFETY, a benchmark for evaluating representation steering methods across nine safety perspectives spanning 18 datasets. While prior work highlights the general capabilities of representation steering, we focus on safety perspectives including refusal, bias, hallucination, social behaviors, reasoning, epistemic integrity, and normative judgment. STEERINGSAFETY provides modularized building blocks for state-of-the-art steering methods, enabling unified implementation of DIM, ACE, CAA, PCA, and LAT with recent enhancements such as conditional steering. Results on Gemma-2-2B, Llama-3.1-8B, and Qwen-2.5-7B show that strong steering performance depends on the pairing of method, model, and specific perspective. For instance, DIM is consistently effective, yet all methods exhibit substantial entanglement, where improving effectiveness on one safety perspective often significantly changes performance on others. Social behaviors are most vulnerable (degradation up to 76%), refusal steering (jailbreaking) frequently compromises normative judgment such as commonsense morality (up to 26%), and hallucination steering shifts political views unpredictably across models, ranging from a 25% shift to the right to a 28% shift to the left. These findings show the need to understand steering methods through multiple safety angles rather than a single target behavior.} }
Endnote
%0 Conference Paper %T SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives %A Vincent Siu %A Nicholas Crispino %A David Park %A Nathan W. Henry %A Zhun Wang %A Yang Liu %A Dawn Song %A Chenguang Wang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-siu26a %I PMLR %P 113832--113884 %U https://proceedings.mlr.press/v306/siu26a.html %V 306 %X We introduce STEERINGSAFETY, a benchmark for evaluating representation steering methods across nine safety perspectives spanning 18 datasets. While prior work highlights the general capabilities of representation steering, we focus on safety perspectives including refusal, bias, hallucination, social behaviors, reasoning, epistemic integrity, and normative judgment. STEERINGSAFETY provides modularized building blocks for state-of-the-art steering methods, enabling unified implementation of DIM, ACE, CAA, PCA, and LAT with recent enhancements such as conditional steering. Results on Gemma-2-2B, Llama-3.1-8B, and Qwen-2.5-7B show that strong steering performance depends on the pairing of method, model, and specific perspective. For instance, DIM is consistently effective, yet all methods exhibit substantial entanglement, where improving effectiveness on one safety perspective often significantly changes performance on others. Social behaviors are most vulnerable (degradation up to 76%), refusal steering (jailbreaking) frequently compromises normative judgment such as commonsense morality (up to 26%), and hallucination steering shifts political views unpredictably across models, ranging from a 25% shift to the right to a 28% shift to the left. These findings show the need to understand steering methods through multiple safety angles rather than a single target behavior.
APA
Siu, V., Crispino, N., Park, D., Henry, N.W., Wang, Z., Liu, Y., Song, D. & Wang, C.. (2026). SteeringSafety: Benchmarking Representation Steering in LLMs Across Safety Perspectives. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:113832-113884 Available from https://proceedings.mlr.press/v306/siu26a.html.

Related Material