Disentangling steering vectors improves control of large language models
Disentangling Steering Vectors
Machine Learning
Summary
Large language models can be nudged to behave in certain ways by changing their internal activations using steering vectors. However, these vectors often mix several ideas or styles together, making it hard to control the model precisely. The authors propose a method that separates these mixed effects into clearer, distinct parts by analyzing differences in model activations and using a special training technique called a sparse autoencoder. Their approach makes it easier to understand and direct the model’s behavior more accurately. Tests on different datasets and models show that this method produces clearer and more predictable control signals.
large language modelsactivation steeringsteering vectorsdisentanglementsparse autoencodermodel interpretabilityinference-time controlsemantic features
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.