How far can the political views of language models be steered?
A paper on how far system-prompt instructions can steer the answers of large language models was presented at AIES 2026 in Malmö.
About the project
Large language models increasingly take part in decision making, while the values and preferences they follow are hard to see and verify. Powerful but misaligned AI systems pose a serious practical risk, and existing methods for understanding and aligning the values of such systems are still limited.
The project Value-aligned and interpretable optimization and reasoning (VALOR) develops methods for interpretability and value alignment of large language models (LLMs) in optimization and reasoning processes. This goal breaks down into:
- methods for better understanding how models make decisions,
- techniques for monitoring and steering model behaviour towards target values,
- open benchmark datasets and tools for evaluating value alignment.
VALOR is a four-year institutional research project of the University of Zagreb Faculty of Electrical Engineering and Computing (FER), running from October 2025 to September 2029. Principal investigator: Mario Brčić.
Research team
- izv. prof. dr. sc. Mario BrčićZPR
- izv. prof. dr. sc. Domagoj VlahZPM
- izv. prof. dr. sc. Mate PuljizZPM
- doc. dr. sc. Stjepan ŠebekZPM
- izv. prof. dr. sc. Vladimir KlemoZEMRIS
- doc. dr. sc. Marko SossichZPF
- dr. sc. Agneza KrajnaZPR
- Mihael Kovač, mag. ing. comp.ZPR
- dr. sc. Kristijan PojeZPR
- Iva Mikuš (Manojlović), mag. math.ZPM
- Luka Hobor, mag. ing. comp.ZPR
News
All news →Visiting researcher Dr. Wojciech Cygan from the University of Wrocław
Dr. Wojciech Cygan (University of Wrocław) visited FER and worked with Stjepan Šebek on fluctuation laws of random walks with drift.
VALOR at the European Frontier AI-Safety Day in Berlin
The principal investigator took part in a meeting of European frontier AI safety researchers.
Better bounds for the "inradius" of Brownian motion
Stjepan Šebek and co-authors improved the bounds on the inradius of the convex hull of planar Brownian motion.
VALOR at the 4th Korean–Croatian Summer Probability Camp in Dubrovnik
Stjepan Šebek took part in the Korean–Croatian summer probability camp in Dubrovnik.
Stable long-horizon prediction of partial differential equation solutions
At Equadiff 2026 in Prague, Domagoj Vlah presented latent-tensor vision transformers for fast and stable prediction of parametric PDE solutions.
Publications
Journal papers
- Reduced order modeling of partial differential equations on parameter-dependent domains using deep neural networks Computers & Mathematics with Applications 216 (2026) 282–300DOIarXiv
Conference papers
- Auditing Alignment Controllability in LLMs via Political Axes AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026), MalmöarXivGitHubZenodo
Preprints
- Improved bounds on the inradius and inverse inradius process of the convex hull of planar Brownian motion arXiv:2607.26944, 2026arXiv
- On the convex hull of a planar Brownian bridge with a random Gaussian endpoint arXiv:2606.24485, 2026arXiv
- Expected perimeter and area of the convex hull of planar Brownian motion stopped upon exiting the unit disk arXiv:2604.18427, 2026arXiv
- AE-ViT: Stable Long-Horizon Parametric Partial Differential Equations Modeling arXiv:2604.06475, 2026arXiv
Talks
- Stable Long-Horizon Prediction of Parametric PDE Solution with Latent-Tensor Vision Transformers Equadiff 2026, Prag / Prague