Measuring Sparse Autoencoder Feature Sensitivity

2025/09/28 by Tian, Claire, Katherine Tian, Nan Hu +2
Computer Science · #Anomaly Detection Techniques and Applications #Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Image Processing and 3D Reconstruction

paper · pdf · doi:10.48550/arxiv.2509.23717

openalex publication_date 2025/09/28 · openalex created_date 2025/10/19 · openalex updated_date 2026/07/28

Abstract

Sparse Autoencoder (SAE) features have become essential tools for mechanistic interpretability research. SAE features are typically characterized by examining their activating examples, which are often "monosemantic" and align with human interpretable concepts. However, these examples don't reveal feature sensitivity: how reliably a feature activates on texts similar to its activating examples. In this work, we develop a scalable method to evaluate feature sensitivity. Our approach avoids the need to generate natural language descriptions for features; instead we use language models to generate text with the same semantic properties as a feature's activating examples. We then test whether the feature activates on these generated texts. We demonstrate that sensitivity measures a new facet of feature quality and find that many interpretable features have poor sensitivity. Human evaluation confirms that when features fail to activate on our generated text, that text genuinely resembles the original activating examples. Lastly, we study feature sensitivity at the SAE level and observe that average feature sensitivity declines with increasing SAE width across 7 SAE variants. Our work establishes feature sensitivity as a new dimension for evaluating both individual features and SAE architectures.

Citations

• Learning Multi-Level Features with Matryoshka Sparse Autoencoders
• SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
• Sparse Autoencoders Do Not Find Canonical Units of Analysis
• Sparse Autoencoders for Hypothesis Generation
• Enhancing Neural Network Interpretability with Feature-Aligned Sparse Autoencoders
• Applying sparse autoencoders to unlearn knowledge in language models
• Automatically Interpreting Millions of Features in Large Language Models
• A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders
• Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2
• Measuring Progress in Dictionary Learning for Language Model Interpretability with Board Game Models
• Transcoders Find Interpretable LLM Feature Circuits
• Scaling and evaluating sparse autoencoders
• Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control
• A Multimodal Automated Interpretability Agent
• RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representations
• Rigorously Assessing Natural Language Explanations of Neurons
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• Explaining black box text modules in natural language with language models
• Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
• The Pile: An 800GB Dataset of Diverse Text for Language Modeling

Related

• Evaluating Adversarial Robustness of Concept Representations in Sparse Autoencoders
• SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
• Scaling and evaluating sparse autoencoders
• Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects
• Do Sparse Autoencoders Generalize? A Case Study of Answerability