mlsanityData to 7 Oct 2026

Sparse Autoencoders Do Not Find Canonical Units of Analysis

arXiv 2502.04878 ↗First posted 7 Feb 2025cs.LG cs.AIcited by 80

Not on the map. Why?

Abstract (arXiv)

A common goal of mechanistic interpretability is to decompose the activations of neural networks into features: interpretable properties of the input computed by the model. Sparse autoencoders (SAEs) are a popular method for finding these features in LLMs, and it has been postulated that they can be used to find a textitcanonical set of units: a unique and complete list of atomic features. We cast doubt on this belief using two novel techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic. SAE stitching involves inserting or swapping latents from a larger SAE into a smaller one. Latents from the larger SAE can be divided into two categories: novel latents, which improve performance when added to the smaller SAE, indicating they capture novel information, and reconstruction latents, which can replace corresponding latents in the smaller SAE that have similar behavior. The existence of novel features indicates incompleteness of smaller SAEs. Using meta-SAEs -- SAEs trained on the decoder matrix of another SAE -- we find that latents in SAEs often decompose into combinations of latents from a smaller SAE, showing that larger SAE latents are not atomic. The resulting decompositions are often interpretable; e.g. a latent representing ``Einstein'' decomposes into ``scientist'', ``Germany'', and ``famous person''. Even if SAEs do not find canonical units of analysis, they may still be useful tools. We suggest that future research should either pursue different approaches for identifying such units, or pragmatically choose the SAE size suited to their task. We provide an interactive dashboard to explore meta-SAEs: https://metasaes.streamlit.app/

Key facts. Cited by 80 papers in the map (data through 7 Oct 2026). First posted on arXiv 7 Feb 2025.

Builds on

0 of 0 references.

No references here yet.

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 32; 2026: 48.

80 papers.

Sort:

By area:

  1. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 136
  2. 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometrycited by 58
  3. 2025Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atomscited by 23
  4. 2025Dense SAE Latents Are Features, Not Bugscited by 17
  5. 2025Feature-Level Insights into Artificial Text Detection with Sparse Autoencoderscited by 14
  6. 2025Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEscited by 14
  7. 2025Stochastic Parameter Decompositioncited by 14
  8. 2026Do Sparse Autoencoders Identify Reasoning Features in Language Models?cited by 11
  9. 2025Measuring and Guiding Monosemanticitycited by 10
  10. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?cited by 10
  11. 2025Does higher interpretability imply better utility? A Pairwise Analysis on Sparse Autoencoderscited by 9
  12. 2025Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoderscited by 8
38 more
  1. 2025Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrievalcited by 8
  2. 2025OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Featurescited by 7
  3. 2025Evaluating and Designing Sparse Autoencoders by Approximating Quasi-Orthogonalitycited by 6
  4. 2025Mechanistic Indicators of Understanding in Large Language Modelscited by 6
  5. 2026From Atoms to Trees: Building a Structured Feature Forest with Hierarchical Sparse Autoencoderscited by 6
  6. 2025Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoderscited by 5
  7. 2025Teach Old SAEs New Domain Tricks with Boostingcited by 5
  8. 2025Measuring Sparse Autoencoder Feature Sensitivitycited by 4
  9. 2025Feature Identification via the Empirical NTKcited by 4
  10. 2025Superposition disentanglement of neural representations reveals hidden alignmentcited by 4
  11. 2025Exploratory Causal Inference in SAEncecited by 4
  12. 2025AlignSAE: Concept-Aligned Sparse Autoencoderscited by 4
  13. 2025Attribution-Guided Distillation of Matryoshka Sparse Autoencoderscited by 4
  14. 2025Ensembling Sparse Autoencoderscited by 3
  15. 2025Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Modelscited by 3
  16. 2025Mechanistic Interpretability with Sparse Autoencoder Neural Operatorscited by 3
  17. 2026Towards Worst-Case Guarantees with Scale-Aware Interpretabilitycited by 3
  18. 2026Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoderscited by 3
  19. 2025Representation Learning on a Random Latticecited by 2
  20. 2025Evaluating Explanations: An Explanatory Virtues Framework for Mechanistic Interpretability -- The Strange Science Part I.iicited by 2
  21. 2025FaithfulSAE: Towards Capturing Faithful Features with Sparse Autoencoders without External Dataset Dependenciescited by 2
  22. 2025Visual Exploration of Feature Relationships in Sparse Autoencoders with Curated Conceptscited by 2
  23. 2026Superposition Without Interference? Towards Isolated Interventions via Almost Orthogonal Features in Language Modelscited by 2
  24. 2026Geometric Routing Enables Causal Expert Control in Mixture of Expertscited by 2
  25. 2026Descriptive Collision in Sparse Autoencoder Auto-Interpretability: When One Explanation Describes Many Featurescited by 2
  26. 2026Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretabilitycited by 2
  27. 2025Towards Open-Ended Visual Scientific Discovery with Sparse Autoencoderscited by 1
  28. 2025Enforcing Orderedness to Improve Feature Consistencycited by 1
  29. 2026Singular Vectors of Attention Heads Align with Featurescited by 1
  30. 2026A Unifying Framework for Unsupervised Concept Extractioncited by 1
  31. 2026Bilinear autoencoders find interpretable manifoldscited by 1
  32. 2026Disentangled Sparse Representations for Concept-Separated Diffusion Unlearningcited by 1
  33. 2026A Geometric View for Understanding Concept Learning and Neuron Interpretation in Sparse Autoencoderscited by 1
  34. 2026Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?cited by 1
  35. 2026Retrieval is Enough: Training-Free Interpretability with a Tool-Using Agentcited by 1
  36. 2026Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effectscited by 1
  37. 2025SplInterp: Improving our Understanding and Training of Sparse Autoencoderscited by 0
  38. 2026Supervised sparse auto-encoders for interpretable and compositional representationscited by 0

50 most cited of 80 All 80 (JSON)