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

• URL: http://arxiv.org/abs/2607.24645v1
• Date: Mon, 27 Jul 2026 16:45:08 GMT
• Title: Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects
• Abstract summary: Feature-Effect Geometry Analysis (FEGA) is an unsupervised framework that removes the same active SAE feature across contexts.<n>We show that a feature can be interpretable and causally relevant without providing a stable direction for steering.
• Score: 58.98080129016082
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects [6.312016976793988]
  We propose emphConditionally Additive Local Models (CALMs)<n>CALMs balance interpretability of GAMs with the accuracy of GA$2$Ms.<n>Experiments show CALMs consistently outperform GAMs and achieve accuracy comparable with GA$2$Ms.
  arXiv  Detail & Related papers  (2026-02-18T14:45:33Z)
• Control Reinforcement Learning: Interpretable Token-Level Steering of LLMs via Sparse Autoencoder Features [1.5874067490843806]
  Control Reinforcement Learning trains a policy to select SAE features for steering at each token, producing interpretable intervention logs.<n> Adaptive Feature Masking encourages diverse feature discovery while preserving singlefeature interpretability.<n>On Gemma 2 2B across MMLU, BBQ, GSM8K, HarmBench, and XSTest, CRL achieves improvements while providing per-token intervention logs.
  arXiv  Detail & Related papers  (2026-02-11T02:28:49Z)
• TangledFeatures: Robust Feature Selection in Highly Correlated Spaces [0.0]
  We introduce TangledFeatures, a framework for feature selection in correlated feature spaces.<n>It identifies representative features from groups of entangled predictors, reducing redundancy while retaining explanatory power.<n>We show that the selected features correspond to structurally meaningful intra-atomic distances that explain variation in backbone torsional angles.
  arXiv  Detail & Related papers  (2025-10-16T05:54:04Z)
• SAEs Are Good for Steering -- If You Select the Right Features [45.47261543304217]
  Current methods identify SAE features to steer by analyzing the input tokens that activate them.<n>In this work, we draw a distinction between two types of features: input features, which mainly capture patterns in the model's input, and output features, which have a human-understandable effect on the model's output.
  arXiv  Detail & Related papers  (2025-05-26T14:47:59Z)
• Going Beyond Neural Network Feature Similarity: The Network Feature Complexity and Its Interpretation Using Category Theory [64.06519549649495]
  We provide the definition of what we call functionally equivalent features. These features produce equivalent output under certain transformations. We propose an efficient algorithm named Iterative Feature Merging.
  arXiv  Detail & Related papers  (2023-10-10T16:27:12Z)
• EEGminer: Discovering Interpretable Features of Brain Activity with Learnable Filters [72.19032452642728]
  We propose a novel differentiable EEG decoding pipeline consisting of learnable filters and a pre-determined feature extraction module. We demonstrate the utility of our model towards emotion recognition from EEG signals on the SEED dataset and on a new EEG dataset of unprecedented size. The discovered features align with previous neuroscience studies and offer new insights, such as marked differences in the functional connectivity profile between left and right temporal areas during music listening.
  arXiv  Detail & Related papers  (2021-10-19T14:22:04Z)
• Can contrastive learning avoid shortcut solutions? [88.249082564465]
  implicit feature modification (IFM) is a method for altering positive and negative samples in order to guide contrastive models towards capturing a wider variety of predictive features. IFM reduces feature suppression, and as a result improves performance on vision and medical imaging tasks.
  arXiv  Detail & Related papers  (2021-06-21T16:22:43Z)
• Unsupervised Feature Learning for Manipulation with Contrastive Domain Randomization [19.474628552656764]
  We show that a naive application of domain randomization to unsupervised learning does not promote invariance. We propose a simple modification of the contrastive loss to fix this, exploiting the fact that we can control the simulated randomization of visual properties.
  arXiv  Detail & Related papers  (2021-03-20T09:54:45Z)
• Removing Spurious Features can Hurt Accuracy and Affect Groups Disproportionately [83.68135652247496]
  A natural remedy is to remove spurious features from the model. We show that removal of spurious features can decrease accuracy due to inductive biases. We also show that robust self-training can remove spurious features without affecting the overall accuracy.
  arXiv  Detail & Related papers  (2020-12-07T23:08:59Z)
• The role of feature space in atomistic learning [62.997667081978825]
  Physically-inspired descriptors play a key role in the application of machine-learning techniques to atomistic simulations. We introduce a framework to compare different sets of descriptors, and different ways of transforming them by means of metrics and kernels. We compare representations built in terms of n-body correlations of the atom density, quantitatively assessing the information loss associated with the use of low-order features.
  arXiv  Detail & Related papers  (2020-09-06T14:12:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.