Comparing Bottom-Up and Top-Down Steering Approaches on In-Context Learning Tasks

Madeline Brumley, Joe Kwon, David Krueger, Dmitrii Krasheninnikov, Usman Anwar

Keywords: LLMs, interpretability, evaluation, steering vectors, activation steering, representation engineering

TL;DR: We evaluate steering vectors computed with top-down and bottom-up methods, and find that both methods have more significant limitations on ICL tasks than previously suggested.

Abstract: A key objective of interpretability research on large language models (LLMs) is to develop methods for robustly steering models toward desired behaviors. To this end, two distinct approaches to interpretability—"bottom-up" and "top-down"—have been presented, but there has been little quantitative comparison between them. We present a case study comparing the effectiveness of representative vector steering methods from each branch: function vectors (FVs) [Todd et al., 2024], as a bottom-up method, and in-context vectors (ICVs) [Liu et al., 2024] as a top-down method. While both aim to capture compact representations of broad in-context learning tasks, we find they are effective only on specific types of tasks: ICVs outperform FVs in behavioral shifting, whereas FVs excel in tasks requiring more precision. We discuss the implications for future evaluations of steering methods and for further research into top-down and bottom-up steering given these findings.

Email Of Author Nominated As Reviewer: [e-mail]

Submission Number: 28