A Unified Understanding and Evaluation of Steering Methods
Shawn Im, Sharon Li
TL;DR
This paper tackles the lack of a unified framework for understanding and evaluating latent space steering in large language models. It formalizes a framework around contrastive datasets and defines steering vectors that shift intermediate activations, proving that the optimal vector is the mean difference $\mathbf{v} = \mathbb{E}[\mathbf{h}_+ - \mathbf{h}_-]$, i.e., the Mean of Differences. Through extensive experiments on multiple-choice QA and open-ended generation, it shows that MoD consistently outperforms PCA-based and classifier-based methods and explains why such methods can misalign direction or scale. The work also identifies the residual stream and layer ~13 as favorable for embedding extraction and provides actionable guidance for designing, optimizing, and deploying steering vectors, while acknowledging data-dependent limitations and future directions for more fine-grained steering. The findings have practical impact for reliably controlling LLM behavior without retraining and offer a principled basis for comparing and improving steering techniques across tasks and datasets.
Abstract
Latent space steering methods provide a practical approach to controlling large language models by applying steering vectors to intermediate activations, guiding outputs toward desired behaviors while avoiding retraining. Despite their growing importance, the field lacks a unified understanding and consistent evaluation across tasks and datasets, hindering progress. This paper introduces a unified framework for analyzing and evaluating steering methods, formalizing their core principles and offering theoretical insights into their effectiveness. Through comprehensive empirical evaluations on multiple-choice and open-ended text generation tasks, we validate these insights, identifying key factors that influence performance and demonstrating the superiority of certain methods. Our work bridges theoretical and practical perspectives, offering actionable guidance for advancing the design, optimization, and deployment of latent space steering methods in LLMs.