Feature reliability drops in scaled sparse autoencoders with active budget
Active Budget Can Kill Sensitivity: Diagnosing and Repairing TopK Sparse Autoencoder Reliability
Artificial Intelligence
Summary
Sometimes computers try to find important features in complex data by picking only a few at a time. The authors showed that when you make these models wider and force them to choose only a limited number of features (an active budget), some rare but meaningful features become less reliable. This happens because of how the model decides which features to keep, especially near a cutoff point. The authors came up with a way to fix this by stabilizing the order in which features are chosen, improving the reliability of rare features without hurting the model’s performance.
What this means in practice
- •For machine learning engineers: Improve interpretation tools by ensuring features extracted remain stable across variations in input data.
- •For language model developers: Develop better sparse representations that keep rare but important features reliable for analysis of large language model activations.
Authors
Zhenting Huang, Junnan Liu, Qianren Mao, Zhixing Tan, Bo Jiang
Abstract
Sparse autoencoders (SAEs) are increasingly scaled to wider dictionaries to recover fine-grained structure from large language model activations. However, a feature is useful for interpretation only if it remains a stable unit of analysis when the same meaning is expressed in different surface forms. We study this reliability question for TopK SAEs via feature sensitivity. Experiments demonstrate that scaling selectively reduces the sensitivity of rare features, while common features remain comparatively stable. A controlled width\(\times k\) factorial experiment identifies the active budget k as the root cause: the degradation arises from the selection boundary rather than dictionary width alone. We attribute this failure to the geometry of TopK selection. The active margin, the distance to the cutoff, predicts feature loss without thresholds. Guided by this margin diagnosis, we introduce pairwise rank stabilization. Our method targets ordering failures at the cutoff and improves rare-feature sensitivity by \(8.83\) percentage points, while keeping reconstruction and alive-feature coverage near the baseline. Overall, our results suggest that wide TopK SAEs should be evaluated not only by reconstruction, sparsity, and feature count, but also by feature reliability under semantic variation and boundary geometry for stable interpretability.