Modular Pretraining Enables Access Control

Ethan Roland, Murat Cubuktepe, Erick Martinez, Stijn Servaes, Keenan Pepper, Michael Vaiana, Diogo S De Lucena, Judd Rosenblatt, Addie Foote, Cem Anil, Alex Cloud
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:105584-105617, 2026.

Abstract

AI developers face a dual-use dilemma. An AI capability that helps one user cure a disease can help another synthesize one. This dilemma could be resolved with access control, limiting dual-use AI capabilities to trusted deployments with a legitimate need. A gold standard for access control would be to serve separate models with different capabilities to different users. However, training and deploying multiple models is prohibitively expensive. To address this challenge, we propose gradient-routed auxiliary modules (GRAM), a pre-training method that adds modules to a neural network and selectively updates them to induce specialization. Ablating a module at inference time removes its capability from the network, approximating a model trained on filtered data. We evaluate GRAM on synthetic stories and realistic dual-use data spanning virology, cybersecurity, nuclear physics, and specialized code. These experiments show that GRAM disables targeted capabilities while preserving the rest, and resists their recovery under finetuning better than post-hoc unlearning. Most importantly, a Chinchilla-optimal scaling analysis from 50M to 5B parameters shows that the gap between data-filtered and full-data models widens with scale on removed capabilities but stays small on retained ones, and that GRAM closely tracks data filtering. GRAM’s training cost is independent of the number of supported capability profiles, yielding a 5$\times$ reduction over data filtering in our 5-profile setting.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-roland26a, title = {Modular Pretraining Enables Access Control}, author = {Roland, Ethan and Cubuktepe, Murat and Martinez, Erick and Servaes, Stijn and Pepper, Keenan and Vaiana, Michael and De Lucena, Diogo S and Rosenblatt, Judd and Foote, Addie and Anil, Cem and Cloud, Alex}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {105584--105617}, 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/roland26a/roland26a.pdf}, url = {https://proceedings.mlr.press/v306/roland26a.html}, abstract = {AI developers face a dual-use dilemma. An AI capability that helps one user cure a disease can help another synthesize one. This dilemma could be resolved with access control, limiting dual-use AI capabilities to trusted deployments with a legitimate need. A gold standard for access control would be to serve separate models with different capabilities to different users. However, training and deploying multiple models is prohibitively expensive. To address this challenge, we propose gradient-routed auxiliary modules (GRAM), a pre-training method that adds modules to a neural network and selectively updates them to induce specialization. Ablating a module at inference time removes its capability from the network, approximating a model trained on filtered data. We evaluate GRAM on synthetic stories and realistic dual-use data spanning virology, cybersecurity, nuclear physics, and specialized code. These experiments show that GRAM disables targeted capabilities while preserving the rest, and resists their recovery under finetuning better than post-hoc unlearning. Most importantly, a Chinchilla-optimal scaling analysis from 50M to 5B parameters shows that the gap between data-filtered and full-data models widens with scale on removed capabilities but stays small on retained ones, and that GRAM closely tracks data filtering. GRAM’s training cost is independent of the number of supported capability profiles, yielding a 5$\times$ reduction over data filtering in our 5-profile setting.} }
Endnote
%0 Conference Paper %T Modular Pretraining Enables Access Control %A Ethan Roland %A Murat Cubuktepe %A Erick Martinez %A Stijn Servaes %A Keenan Pepper %A Michael Vaiana %A Diogo S De Lucena %A Judd Rosenblatt %A Addie Foote %A Cem Anil %A Alex Cloud %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-roland26a %I PMLR %P 105584--105617 %U https://proceedings.mlr.press/v306/roland26a.html %V 306 %X AI developers face a dual-use dilemma. An AI capability that helps one user cure a disease can help another synthesize one. This dilemma could be resolved with access control, limiting dual-use AI capabilities to trusted deployments with a legitimate need. A gold standard for access control would be to serve separate models with different capabilities to different users. However, training and deploying multiple models is prohibitively expensive. To address this challenge, we propose gradient-routed auxiliary modules (GRAM), a pre-training method that adds modules to a neural network and selectively updates them to induce specialization. Ablating a module at inference time removes its capability from the network, approximating a model trained on filtered data. We evaluate GRAM on synthetic stories and realistic dual-use data spanning virology, cybersecurity, nuclear physics, and specialized code. These experiments show that GRAM disables targeted capabilities while preserving the rest, and resists their recovery under finetuning better than post-hoc unlearning. Most importantly, a Chinchilla-optimal scaling analysis from 50M to 5B parameters shows that the gap between data-filtered and full-data models widens with scale on removed capabilities but stays small on retained ones, and that GRAM closely tracks data filtering. GRAM’s training cost is independent of the number of supported capability profiles, yielding a 5$\times$ reduction over data filtering in our 5-profile setting.
APA
Roland, E., Cubuktepe, M., Martinez, E., Servaes, S., Pepper, K., Vaiana, M., De Lucena, D.S., Rosenblatt, J., Foote, A., Anil, C. & Cloud, A.. (2026). Modular Pretraining Enables Access Control. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:105584-105617 Available from https://proceedings.mlr.press/v306/roland26a.html.

Related Material