Evaluating Open-Source Sparse Autoencoders on Disentangling Factual Knowledge in GPT-2 Small

TL;DR

This paper evaluates the effectiveness of sparse autoencoders in disentangling factual knowledge in GPT-2 small, finding they underperform compared to neurons and linear features, thus questioning their utility for causal interpretability.

Contribution

It provides the first systematic evaluation of open-source sparse autoencoders on GPT-2 small for causal disentanglement using the RAVEL benchmark.

Findings

SAEs underperform neuron baseline in disentangling factual knowledge

SAEs do not approach the performance of linear features learned via DAS

Code for the evaluation is publicly released

Abstract

A popular new method in mechanistic interpretability is to train high-dimensional sparse autoencoders (SAEs) on neuron activations and use SAE features as the atomic units of analysis. However, the body of evidence on whether SAE feature spaces are useful for causal analysis is underdeveloped. In this work, we use the RAVEL benchmark to evaluate whether SAEs trained on hidden representations of GPT-2 small have sets of features that separately mediate knowledge of which country a city is in and which continent it is in. We evaluate four open-source SAEs for GPT-2 small against each other, with neurons serving as a baseline, and linear features learned via distributed alignment search (DAS) serving as a skyline. For each, we learn a binary mask to select features that will be patched to change the country of a city without changing the continent, or vice versa. Our results show that SAEs struggle to reach the neuron baseline, and none come close to the DAS skyline. We release code here: https://github.com/MaheepChaudhary/SAE-Ravel