Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

2026-07-27

AI Summary

This work addresses the limited interpretability utility of sparse autoencoder (SAE) features due to their unstable causal influence on model behavior. To systematically analyze the downstream effects of feature interventions, the authors propose the Feature Effect Geometric Analysis (FEGA) framework, which—through cross-context ablation of identical SAE features and modeling of their effect geometry from the perspective of output logit changes—reveals that most SAE features lack consistent one-dimensional effects. Integrating unsupervised ablation, geometric analysis, and feature categorization, FEGA demonstrates that value-like features exhibit low-dimensional yet multidirectional impacts, whereas pointer-like features display more diffuse effects. The study further shows that interpretability does not necessarily entail stable, controllable directions of influence.

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.

Problem

• sparse autoencoders
• interpretability
• feature effects
• causal effects
• model steering

Innovation

• Sparse Autoencoders
• Feature-Effect Geometry Analysis
• Interpretability
• Causal Effects
• Value-like vs Pointer-like Features