Toward universal steering and monitoring of AI models
Science · 19.02.2026 · 391 (6787) · 787-792
DOI: 10.1126/science.aea6792 ↗Kurz gefasst
Fachartikel in der Zeitschrift Science, erschienen am 19.02.2026. Verfasst von Daniel Beaglehole und 3 weiteren Autor:innen. Beteiligte Einrichtung: Department of Computer Science and Engineering, University of California San Diego, La Jolla, CA, USA. Die Arbeit wurde bisher 6-mal zitiert (Crossref) und nennt 49 Literaturangaben. Gefördert von: National Science Foundation, Office of Naval Research, Defense Advanced Research Projects Agency.
Abstract (Originalsprache)
Artificial intelligence (AI) models contain much of human knowledge. Understanding the representation of this knowledge will lead to improvements in model capabilities and safeguards. Building on advances in feature learning, we developed an approach for extracting linear representations of semantic notions or concepts in AI models. We showed how these representations enabled model steering, through which we exposed vulnerabilities and improved model capabilities. We demonstrated that concept representations were transferable across languages and enabled multiconcept steering. Across hundreds of concepts, we found that larger models were more steerable and that steering improved model capabilities beyond prompting. We showed that concept representations were more effective for monitoring misaligned content than for using judge models. Our results illustrate the power of internal representations for advancing AI safety and model capabilities.
Autor:innen und Institutionen
- Department of Computer Science and Engineering, University of California San Diego, La Jolla, CA, USA.
- Massachusetts Institute of Technology, Cambridge, MA, USA.
- Broad Institute of MIT and Harvard, Cambridge, MA, USA.
- Wharton School, University of Pennsylvania, Philadelphia, PA, USA.
- Halıcıoğlu Data Science Institute, University of California San Diego, La Jolla, CA, USA.
ORCID-Profile: Daniel Beaglehole, Adityanarayanan Radhakrishnan, Enric Boix-Adserà, Mikhail Belkin
Förderung
- National Science Foundation
- Office of Naval Research
- Defense Advanced Research Projects Agency
- Simons Foundation
- National Science Foundation - Mathematical Foundations of Artificial Intelligence
Zitieren
APA
Beaglehole, D., Radhakrishnan, A., Boix-Adserà, E., & Belkin, M. (2026). Toward universal steering and monitoring of AI models. Science, 391(6787), 787-792. https://doi.org/10.1126/science.aea6792
Vancouver
Beaglehole D, Radhakrishnan A, Boix-Adserà E, Belkin M. Toward universal steering and monitoring of AI models. Science. 2026;391(6787):787-792. doi:10.1126/science.aea6792
BibTeX
@article{Beaglehole2026,
author = {Beaglehole, Daniel and Radhakrishnan, Adityanarayanan and Boix-Adserà, Enric and Belkin, Mikhail},
title = {Toward universal steering and monitoring of AI models},
journal = {Science},
year = {2026},
volume = {391},
number = {6787},
pages = {787--792},
doi = {10.1126/science.aea6792}
}
EUVIMED