Specifying steering targets and decision authority in pluralistic alignment

Determine principled procedures for specifying the target behaviors that large language model steering should aim to achieve in value‑pluralistic settings and establish who should be responsible for deciding these steering targets.

Background

The paper discusses ethical considerations of providing tools for steering generative models, noting that steerability is considered useful for value‑pluralistic systems. However, the authors point out a lack of consensus on how steering targets should be specified and who should make these decisions.

Clarifying both the methodology for specifying targets and the governance of these decisions is necessary to ensure responsible and transparent deployment of steering in real-world settings.