SAEs Are Good for Steering -- If You Select the Right Features

TL;DR

This work distinguishes two roles for Sparse Autoencoder features in language models: input features, which respond to input tokens, and output features, which causally steer generated text. By defining input and output scores through logit-lens projections and counterfactual interventions, the authors show that high-output features are more effective for steering, while high-input features alone are not predictive of steering success. A practical feature-selection method based on output scores yields 2–3x improvements on SAE steering, narrowing the gap with supervised techniques and performing well on AxBench. The findings illuminate layer-wise specialization and offer a principled approach to selecting SAE features for more reliable, efficient steering. This has implications for controllable generation, bias mitigation, and safe deployment of steering techniques in large language models.

Abstract

Sparse Autoencoders (SAEs) have been proposed as an unsupervised approach to learn a decomposition of a model's latent space. This enables useful applications such as steering - influencing the output of a model towards a desired concept - without requiring labeled data. Current methods identify SAE features to steer by analyzing the input tokens that activate them. However, recent work has highlighted that activations alone do not fully describe the effect of a feature on the model's output. In this work, we draw a distinction between two types of features: input features, which mainly capture patterns in the model's input, and output features, which have a human-understandable effect on the model's output. We propose input and output scores to characterize and locate these types of features, and show that high values for both scores rarely co-occur in the same features. These findings have practical implications: after filtering out features with low output scores, we obtain 2-3x improvements when steering with SAEs, making them competitive with supervised methods.

Figures (10)

• Figure 1: Selecting features for steering. (1) Given a concept to steer ("apple"), we amplify a candidate SAE feature during a single forward pass of the model on a neutral prompt. (2) We compute the feature's output score based on the rank and probability of representative after intervention. (3) Features with high output scores are more likely to be effective for steering.
• Figure 2: Examples of steering with input and output features.(a) An input feature, which activates strongly on tokens like "_primary" (leading to a high input score of $0.82$), fails to steer generation meaningfully; with a high steering factor, the model degenerates into repeating the token "school", as if continuing from the word "primary". (b) An output feature, with an output score of $0.81$, yields meaningful, coherent generations when steered at an optimal steering factor.
• Figure 3: Input and output scores across layers in Gemma-2-2B and Gemma-2-9B. The solid lines represent the median input score (blue) and output score (magenta), while the shaded regions denote the interquartile range (25th to 75th percentile), capturing the variability across features within each layer. Early layers are characterized by features with high input scores, while high output scores emerge in later layers.
• Figure 4: Magenta indicates the mean generation success@20 when filtering out features with output scores below different thresholds. Green indicates the mean generation success@20 after filtering randomly sampled sets of features of the same size. Filtering results in significant increase in generation success.
• Figure 5: Even in later layers of the model (16--25 for Gemma-2-2B and 24--41 for Gemma-2-9B), filtering features with low output scores increases mean generation success. Magenta: filtering by output scores. Green: filtering random sets of features of the same size.