analiza aiqbench
This research introduces a parameter-efficient method using Riemannian optimization to steer Large Language Model activations, aiming to improve control over refusal behaviors without relying on external auxiliary vectors. By leveraging Stiefel-constrained rotation steering, the study demonstrates superior intervention efficiency compared to existing techniques that depend on predefined refusal vectors. The work advances the geometric understanding of activation spaces, offering a more reliable framework for aligning model outputs at inference time.
Controlling Refusal Behavior of LLMs via Stiefel-Constrained Rotation Steering
arXiv:2608.30986v1 Announce Type: cross Abstract: Activation steering has emerged as a lightweight approach for controlling model refusal at inference time. A growing line of research explores trainable rotations of activations to develop geometrically principled intervention mechanisms. However, existing techniques rely on auxiliary constructs, such as refusal vectors, to define these rotations. In our work, we develop a self-contained methodology for learning parameter-efficient rotational transformations based on Riemannian optimization. We empirically validate the proposed scheme, demonstrating its superiority in intervention efficiency. An extensive ablation study highlights the importance of key design choices in our method. Our results identify the proposed rotation-based steering scheme as a promising direction for more reliable control over the behavior of LLMs.
Abstract:Activation steering has emerged as a lightweight approach for controlling model refusal at inference time. A growing line of research explores trainable rotations of activations to develop geometrically principled intervention mechanisms. However, existing techniques rely on auxiliary constructs, such as refusal vectors, to define these rotations. In our work, we develop a self-contained methodology for learning parameter-efficient rotational transformations based on Riemannian optimization. We empirically validate the proposed scheme, demonstrating its superiority in intervention efficiency. An extensive ablation study highlights the importance of key design choices in our method. Our results identify the proposed rotation-based steering scheme as a promising direction for more reliable control over the behavior of LLMs.
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2608.30986 [cs.LG]
(or arXiv:2608.30986v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2608.30986 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 208 cuvinte
Subiecte conexe
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · 18 sept. 2026
Game-Theoretic Disempowerment: Loss of Control to Agencyless AI
LessWrong · 27 sept. 2026
SafeTune: A Unified Faithful Library for Auditing and Repairing Safety Drift in Fine-Tuned LLMs
arXiv · 22 sept. 2026
Game Design Creativity evals
LessWrong · 12 oct. 2026