Enhancing SAE-based Steering via Neighbor Integrated Feature Selection
Artificial Intelligence
Summary
The authors study how sparse autoencoders (SAEs) pick features to control large language models better. They find that the common way of choosing features based on top statistical scores is not always the best because important features can have lower scores but still be useful. To fix this, they suggest a new method called Neighbor Integrated Feature Selection (NIFS), which looks at how similar features are to each other when picking them. Testing NIFS on various tasks shows it works better than just picking the top-scoring features.
Authors
Yutian Liu, Xu Wang, Difan Zou
Abstract
Sparse autoencoders (SAEs) disentangle model activations into interpretable features and are widely used for steering large language models. Most existing SAE-based steering methods select features by applying a top- filter based on statistical scores, assuming that higher-scoring features yield stronger steering effects. In this paper, we show that this assumption is often invalid, leading to suboptimal feature selection. Our analysis reveals that effective steering features may be distributed among representationally adjacent, semantically similar groups induced by feature splitting in SAEs. Within such groups, features may exhibit disparate statistical scores despite having comparable steering influence, causing score-based selection to overlook important features. Based on these observations, we propose \textsc{Neighbor Integrated Feature Selection} (\textsc{NIFS}), a plug-and-play strategy that leverages representation similarity to improve feature selection for steering. We evaluate \textsc{NIFS} across multiple SAE-based steering methods and tasks, and demonstrate consistent performance gains over conventional top-$k$ selection.