NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
A Geometric Account of Activation Steering through Angle-Norm Decomposition
TL;DR
Language models often struggle with controlling behavior through traditional methods. This study introduces a geometric approach called Angle-Norm Decomposition to better understand activation steering.
✦ Why It Matters
Engineers can enhance language model control by using distinct angular and radial parameters for activation steering.
Key Takeaways
- Explore geometric methods for activation steering in language models.
- Utilize distinct angular and radial parameters for better control.
- Consider the hidden-state norm's impact on model stability.
- Conduct empirical studies to validate steering methods in practice.
Full Summary
Controlling the behavior of language models has been challenging, particularly with linear activation steering methods. Recent spherical steering paradigms aim to improve this by addressing the limitations of additive interventions, based on the assumption that the hidden-state norm does not convey relevant information about concepts.
This research revisits that assumption through a controlled empirical study, focusing on the angular and radial components of activation steering. The study involved seven different language models and revealed that concepts are primarily represented in angular structure, supporting the use of spherical methods.
However, the hidden-state norm was found to be important for the stability and downstream effects of steering interventions. These findings suggest that activation steering should be parameterized using distinct angular and radial components, rather than a single additive coefficient that conflates these effects.
This approach can lead to more interpretable and effective steering methods.