mlsanityData to 14 Oct 2026

Position: Use Sparse Autoencoders to Discover Unknowns

arXiv 2506.23845 ↗First posted 30 Jun 2025cs.LG cs.AI cs.CLcited by 19

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

Abstract (arXiv)

While sparse autoencoders (SAEs) have generated significant excitement, a series of negative results have added to skepticism about their usefulness. Here, we establish a conceptual distinction that reconciles competing narratives surrounding SAEs. We argue that even if SAEs may be less effective for textitacting on known concepts, SAEs are especially powerful tools for textitdiscovering unknown concepts. This distinction separates existing negative results from positive results, and suggests several classes of SAE applications. Specifically, we outline use cases for SAEs in (i) ML interpretability, explainability, fairness, auditing, and safety, and (ii) social and health sciences.

Key facts. Cited by 19 papers in the map (data through 14 Oct 2026). Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability. Builds on 29 of its 66 references that are in the map. First posted on arXiv 30 Jun 2025.

Builds on

29 of 66 references.

Closest ancestors · 3

Same direction, newest first

  1. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoders21 Mar
  2. 2025Transcoders Beat Sparse Autoencoders for Interpretability31 Jan
  3. 2022Toy Models of Superposition21 Sep

Other references · 26

  1. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  2. 2020Concept Bottleneck Modelscited by 1,032
  3. 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
  4. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
  5. 2017Network Dissection: Quantifying Interpretability of Deep Visual Representationscited by 708
  6. 2024Scaling and evaluating sparse autoencoderscited by 553
  7. 2019ClinicalBERT: Modeling Clinical Notes and Predicting Hospital Readmissioncited by 493
  8. 2019Linguistic Knowledge and Transferability of Contextual Representationscited by 383
  9. 2018Confounding variables can degrade generalization performance of radiological deep learning modelscited by 379
  10. 2017Word Embeddings Quantify 100 Years of Gender and Ethnic Stereotypescited by 253
  11. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 218
  12. 2013k-Sparse Autoencoderscited by 215
14 more
  1. 2025Open Problems in Mechanistic Interpretabilitycited by 179
  2. 2021Reading Race: AI Recognises Patient's Racial Identity In Medical Imagescited by 156
  3. 2022Natural Language Descriptions of Deep Visual Featurescited by 137
  4. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 115
  5. 2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 80
  6. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
  7. 2024InterPLM: Discovering Interpretable Features in Protein Language Models via Sparse Autoencoderscited by 66
  8. 2024Applying sparse autoencoders to unlearn knowledge in language modelscited by 59
  9. 2025Sparse Autoencoders for Hypothesis Generationcited by 26
  10. 2024Disentangling Dense Embeddings with Sparse Autoencoderscited by 23
  11. 2023Interpretable-by-Design Text Understanding with Iteratively Generated Concept Bottleneckcited by 22
  12. 2019Measuring the Completeness of Theoriescited by 6
  13. 2025BehaviorBox: Automated Discovery of Fine-Grained Performance Differences Between Language Modelscited by 5
  14. 2025Aggregated Individual Reporting for Post-Deployment Evaluationcited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 10; 2026: 9.

19 papers.

Sort:

By area:

  1. 2025What's In My Human Feedback? Learning Interpretable Descriptions of Preference Datacited by 26
  2. 2025AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Featurescited by 9
  3. 2026Interpretable Debiasing of Vision-Language Models for Social Fairnesscited by 6
  4. 2025ActivationReasoning: Logical Reasoning in Latent Activation Spacescited by 5
  5. 2025Person-Centric Annotations of LAION-400M: Auditing Bias and Its Transfer to Modelscited by 4
  6. 2025Exploratory Causal Inference in SAEncecited by 4
  7. 2025Sparse Autoencoders are Topic Modelscited by 3
  8. 2026Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancycited by 3
  9. 2025Re-envisioning Euclid Galaxy Morphology: Identifying and Interpreting Features with Sparse Autoencoderscited by 2
  10. 2025Can SAEs reveal and mitigate racial biases of LLMs in healthcare?cited by 2
  11. 2025Towards Open-Ended Visual Scientific Discovery with Sparse Autoencoderscited by 1
  12. 2025Graph-Regularized Sparse Autoencoders for LLM Safety Steeringcited by 1
7 more
  1. 2026Automated Concept Discovery for LLM-as-a-Judge Preference Analysiscited by 1
  2. 2026Causal Effect Estimation with Latent Textual Treatmentscited by 0
  3. 2026Conditional Hypothesis Generation for LLM-Based Text Analysis with Researcher-Specified Covariatescited by 0
  4. 2026An Investigation of the NeurIPS and ICML 2025 Position Trackscited by 0
  5. 2026Feature Evolution and Migration during Vision Transformer Trainingcited by 0
  6. 2026A Testable Theory of Atomic Featurescited by 0
  7. 2026Sparse Autoencoders for Mechanistic Interpretability in NLP: A Critical Survey of Recoverability, Interpretability, and Utilitycited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Evaluating SAE interpretability without explanations
  2. 2024Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks
  3. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
  4. 2026Verifying the Linear Representation Hypothesis: How Interpretable Are Vision SAEs?
  5. 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry
  6. 2025Sparse Autoencoders Trained on the Same Data Learn Different Features
  7. 2026Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?
  8. 2025SplInterp: Improving our Understanding and Training of Sparse Autoencoders