mlsanityData to 14 Oct 2026

What Do Sparse Autoencoder Feature Discovery Scores Identify? From Activations to Interventions

arXiv 2610.15850 ↗First posted 11 Oct 2026cs.LG cs.CL stat.MLcited by 0

On the map: Mechanistic interpretability → Sparse autoencoder training and evaluation

Abstract (arXiv)

Sparse autoencoder (SAE) features are typically discovered from how their activations co-vary with behaviour, yet they are often used through interventions such as steering or ablation. We take intervention attribution, the first-order effect of a specified intervention on the model output, as our estimand. For a broad class of response functions, we show that observational feature-discovery evidence generally does not identify this attribution, and derive an exact gradient-based decomposition of the gap between linear-probe predictions and intervention attribution. Across four model-SAE pairs, median probe-to-attribution magnitude ratios range from tens to hundreds, while attributions closely track small-step steering. For single-latent ablations with nonzero effects, gradient predictions match the realized sign in 91-97% of cases, compared with 17-35% for probe-implied changes. Attribution-selected sets also produce larger absolute mean ablation effects on held-out prompts than sets selected by correlation, rank tests or probe coefficients. Together, these results establish intervention attribution as a principled, gradient-estimable target for selecting SAE features for intervention.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability. Builds on 30 of its 55 references that are in the map. First posted on arXiv 11 Oct 2026.

Builds on

30 of 55 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025Which Sparse Autoencoder Features Are Real? Model-X Knockoffs for False Discovery Rate Control12 Nov
  2. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 29 Aug

Other references · 28

  1. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  2. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  3. 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
  4. 2019Designing and Interpreting Probes with Control Taskscited by 640
  5. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  6. 2021Causal Abstractions of Neural Networkscited by 422
  7. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 218
  8. 2024How to use and interpret activation patchingcited by 214
  9. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 209
  10. 2026Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnetcited by 191
  11. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 160
  12. 2024Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanismscited by 149
16 more
  1. 2016Panning for Gold: Model-X Knockoffs for High-dimensional Controlled Variable Selectioncited by 121
  2. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 115
  3. 2024AtP*: An efficient and scalable method for localizing LLM behaviour to componentscited by 111
  4. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 92
  5. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
  6. 2020Fairwashing Explanations with Off-Manifold Detergentcited by 72
  7. 2025CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Featurescited by 12
  8. 2026Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoderscited by 6
  9. 2025Mechanistic Interpretability of Code Correctness in LLMs via Sparse Autoencoderscited by 5
  10. 2025Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoderscited by 4
  11. 2026Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretabilitycited by 4
  12. 2025Interpretable Discovery from Unstructured Data: A High-Dimensional Approachcited by 3
  13. 2026When Attribution Patching Lies: Diagnosis and a Second-Order Correctioncited by 3
  14. 2025SAE-FiRE: Enhancing Earnings Surprise Predictions Through Sparse Autoencoder Feature Selectioncited by 2
  15. 2025Disentangled Feature Importancecited by 2
  16. 2026Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decompositioncited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects
  2. 2026Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects
  3. 2026SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behavior
  4. 2026Gradients for Interventions and Activations for Detection: Targeted Feature Learning in Language Models
  5. 2026From Input to Output: A Flexible Agent for Dual-End Interpretation of Sparse Autoencoder Features
  6. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
  7. 2026Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features
  8. 2024Interpreting Attention Layer Outputs with Sparse Autoencoders