Sparse Autoencoder Features for Classifications and Transferability

TL;DR

This paper explores the use of Sparse Autoencoders to extract interpretable features from Large Language Models, demonstrating their effectiveness and transferability across tasks and models, with implications for transparent AI deployment.

Contribution

It systematically analyzes SAE configurations for feature extraction from LLMs, establishing new best practices for interpretability and transferability in safety-critical applications.

Findings

SAE features outperform baselines with macro F1 > 0.8

Features transfer across models from Gemma 2 2B to 9B-IT

Binarization maintains or improves performance while simplifying feature selection

Abstract

Sparse Autoencoders (SAEs) provide potentials for uncovering structured, human-interpretable representations in Large Language Models (LLMs), making them a crucial tool for transparent and controllable AI systems. We systematically analyze SAE for interpretable feature extraction from LLMs in safety-critical classification tasks. Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations. SAE-derived features achieve macro F1 > 0.8, outperforming hidden-state and BoW baselines while demonstrating cross-model transfer from Gemma 2 2B to 9B-IT models. These features generalize in a zero-shot manner to cross-lingual toxicity detection and visual classification tasks. Our analysis highlights the significant impact of pooling strategies and binarization thresholds, showing that binarization offers an efficient alternative to traditional feature selection while maintaining or improving performance. These findings establish new best practices for SAE-based interpretability and enable scalable, transparent deployment of LLMs in real-world applications. Full repo: https://github.com/shan23chen/MOSAIC.