MidSteer: Optimal Affine Framework for Steering Generative Models

Tatiana Gaintseva, Andrew Stepanov, Ziquan Liu, Martin Benning, Gregory Slabaugh, Jiankang Deng, Ismail Elezi
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:32675-32741, 2026.

Abstract

Steering intermediate representations has emerged as a powerful strategy for controlling generative models. However, despite its empirical success, it currently lacks a comprehensive theoretical framework. In this paper, we bridge this gap by formalizing the theory of concept steering. First, we establish a link between steering and affine concept erasure, proving that the standard approach for removing unwanted behaviors is a special case of LEACE (a closed-form method for affine erasure). Next, we formulate a principled theoretical framework for concept switching, LEACE-Switch, and characterize the assumptions under which it provides an optimal affine solution. Building on this analysis, we then introduce MidSteer (Minimal Disturbance concept Steering), a more general affine framework for concept manipulation that relaxes these assumptions and enables directed, minimal-disturbance transformations. We empirically demonstrate that MidSteer performs favorably across a range of tasks, modalities, and architectures, including vision diffusion models and large language models.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-gaintseva26a, title = {{M}id{S}teer: Optimal Affine Framework for Steering Generative Models}, author = {Gaintseva, Tatiana and Stepanov, Andrew and Liu, Ziquan and Benning, Martin and Slabaugh, Gregory and Deng, Jiankang and Elezi, Ismail}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {32675--32741}, 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/gaintseva26a/gaintseva26a.pdf}, url = {https://proceedings.mlr.press/v306/gaintseva26a.html}, abstract = {Steering intermediate representations has emerged as a powerful strategy for controlling generative models. However, despite its empirical success, it currently lacks a comprehensive theoretical framework. In this paper, we bridge this gap by formalizing the theory of concept steering. First, we establish a link between steering and affine concept erasure, proving that the standard approach for removing unwanted behaviors is a special case of LEACE (a closed-form method for affine erasure). Next, we formulate a principled theoretical framework for concept switching, LEACE-Switch, and characterize the assumptions under which it provides an optimal affine solution. Building on this analysis, we then introduce MidSteer (Minimal Disturbance concept Steering), a more general affine framework for concept manipulation that relaxes these assumptions and enables directed, minimal-disturbance transformations. We empirically demonstrate that MidSteer performs favorably across a range of tasks, modalities, and architectures, including vision diffusion models and large language models.} }
Endnote
%0 Conference Paper %T MidSteer: Optimal Affine Framework for Steering Generative Models %A Tatiana Gaintseva %A Andrew Stepanov %A Ziquan Liu %A Martin Benning %A Gregory Slabaugh %A Jiankang Deng %A Ismail Elezi %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-gaintseva26a %I PMLR %P 32675--32741 %U https://proceedings.mlr.press/v306/gaintseva26a.html %V 306 %X Steering intermediate representations has emerged as a powerful strategy for controlling generative models. However, despite its empirical success, it currently lacks a comprehensive theoretical framework. In this paper, we bridge this gap by formalizing the theory of concept steering. First, we establish a link between steering and affine concept erasure, proving that the standard approach for removing unwanted behaviors is a special case of LEACE (a closed-form method for affine erasure). Next, we formulate a principled theoretical framework for concept switching, LEACE-Switch, and characterize the assumptions under which it provides an optimal affine solution. Building on this analysis, we then introduce MidSteer (Minimal Disturbance concept Steering), a more general affine framework for concept manipulation that relaxes these assumptions and enables directed, minimal-disturbance transformations. We empirically demonstrate that MidSteer performs favorably across a range of tasks, modalities, and architectures, including vision diffusion models and large language models.
APA
Gaintseva, T., Stepanov, A., Liu, Z., Benning, M., Slabaugh, G., Deng, J. & Elezi, I.. (2026). MidSteer: Optimal Affine Framework for Steering Generative Models. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:32675-32741 Available from https://proceedings.mlr.press/v306/gaintseva26a.html.

Related Material