Steering vectors disentangled for clearer large language model control
Disentangling Steering Vectors
Machine Learning
Summary
Large language models can be nudged to behave differently by changing what’s happening inside their layers, but often these nudges mix many ideas together, making the changes unpredictable. This paper introduces a way to separate these mixed nudges into clear, individual parts so each one controls a single idea. The authors do this by looking at pairs of different activations and training a special model to pull apart these combined directions. This makes it easier and more precise to guide how language models respond.
What this means in practice
- •For ai system developers: Implement fine-grained control of language model behaviors by isolating distinct steering features during inference.
- •For natural language processing engineers: Design more predictable model adjustments by using disentangled steering directions for specific semantic or stylistic effects.
Authors
Takeru Hiramatsu, Kyohei Atarashi, Koh Takeuchi, Hisashi Kashima
Abstract
Activation steering has emerged as a lightweight, inference-time approach to control the behavior of Large Language Models (LLMs). However, traditional steering vectors used to intervene in LLMs' activations, such as those derived from the difference-in-means method, tend to entangle multiple semantic and stylistic concepts into a single composite direction, leading to unpredictable steering effects. Our core objective is to disentangle this composite direction into its constituent concepts. To this end, we propose Steering Vector Dissection, a framework to explicitly isolate individual and semantically consistent features from these composite directions. Specifically, we pair positive and negative activations and take their differences to generate a set of instance-level steering vectors, and train a dedicated Sparse Autoencoder (SAE) directly on them. Quantitative evaluations across two datasets, two models, and two intervention depths show that our method yields a set of semantically consistent basis vectors whose steering effects are mutually distinguishable. Furthermore, we show that this disentanglement enables precise control over model behaviors.