Ensembling Sparse Autoencoders

Sparse autoencoders (SAEs) are used to decompose neural network activations into human-interpretable features. Typically, features learned by a single SAE are used for downstream applications. However, it has recently been shown that a single SAE captures only a limited subset of features that can be extracted from the activation space. Motivated by this limitation, we introduce and formalize SAE ensembles. Furthermore, we propose to ensemble multiple SAEs through naive bagging and boosting. In naive bagging, SAEs trained with different weight initializations are ensembled, whereas in boosting SAEs sequentially trained to minimize the residual error are ensembled. Theoretically, naive bagging and boosting are justified as approaches to reduce reconstruction error. Empirically, we evaluate our ensemble approaches with three settings of language models and SAE architectures. Our empirical results demonstrate that, compared to an expanded SAE that matches the number of features in the ensemble, ensembling SAEs improves the reconstruction of language model activations along with SAE stability. Additionally, on downstream tasks such as concept detection and spurious correlation removal, SAE ensembles achieve better performance, showing improved practical utility.

Understanding the behavior of Large Language Models is critical given their widespread applications, but these models often learn internal patterns that are difficult for people to understand. Sparse autoencoders are tools that help break these internal representations into simpler, more interpretable pieces, making it easier to study what an LLM has learned and how it behaves. However, a single sparse autoencoder may only capture part of the useful information available inside a model. In this paper, we propose combining multiple sparse autoencoders to reveal additional patterns learned by these models. We propose two ways to combine them: one method, called bagging, averages several independently trained sparse autoencoders, while the other, called boosting, trains them one after another so that each new one focuses on what the previous ones missed. Across several LLMs, we find that these combined approaches often produce more stable and useful features. We also show that they can improve practical tasks such as detecting concepts in text and reducing unwanted gender-related biases.