Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects
2026-07-27 • Machine Learning
Machine LearningArtificial IntelligenceComputation and Language
AI summaryⓘ
The authors studied how features inside sparse autoencoders (SAEs) relate to changes in a model's output, focusing on the geometry of these changes rather than just the features themselves. They developed a method called Feature-Effect Geometry Analysis (FEGA) that looks at how removing certain features affects the model's logits. They found that most features do not create simple, consistent directions for controlling the model; instead, some features represent stable facts while others depend heavily on context. Their work shows that features can be understandable and influence the model without always allowing straightforward control.
sparse autoencoderinterpretabilitylogitsfeature interventionfeature geometryFeature-Effect Geometry Analysis (FEGA)value-like featurespointer-like featuresmodel steering
Authors
Phu Gia Hoang, Anwoy Chatterjee, Tanmoy Chakraborty, Iryna Gurevych, Subhabrata Dutta
Abstract
The wide-scale use of sparse autoencoders (SAEs) as interpretability tools is limited by inconsistent links between SAE features and model behavior. Features with clear activation descriptions may have weak or unexpected causal effects; steering can vary across prompts or oppose the intended direction; and activation-based feature selection can miss features that produce the desired output change. Prior work has studied feature geometry inside the model, where features are computed. We instead study the geometry of changes in model logits caused by feature interventions. We introduce Feature-Effect Geometry Analysis (FEGA), an unsupervised framework that removes the same active SAE feature across contexts and analyzes the resulting cloud of logit changes. Across SAE variants, consistent one-dimensional effects are rare: few features behave like reusable directions. To interpret this variation, we distinguish value-like features, tied to static information such as factual attributes, from pointer-like features, associated with context-dependent operations. Value-like features more often exhibit structured, low-dimensional effects, although these effects typically span several directions. Pointer-like features, by contrast, predominantly exhibit diffuse effects. Our results show that a feature can be interpretable and causally relevant without providing a stable direction for steering.