mlsanityData to 14 Oct 2026

BatchTopK Sparse Autoencoders

arXiv 2412.06410 ↗First posted 9 Dec 2024cs.LG cs.AI stat.MLcited by 130

On the map: Mechanistic interpretability → SAE features and their applications

Abstract (arXiv)

Sparse autoencoders (SAEs) have emerged as a powerful tool for interpreting language model activations by decomposing them into sparse, interpretable features. A popular approach is the TopK SAE, that uses a fixed number of the most active latents per sample to reconstruct the model activations. We introduce BatchTopK SAEs, a training method that improves upon TopK SAEs by relaxing the top-k constraint to the batch-level, allowing for a variable number of latents to be active per sample. As a result, BatchTopK adaptively allocates more or fewer latents depending on the sample, improving reconstruction without sacrificing average sparsity. We show that BatchTopK SAEs consistently outperform TopK SAEs in reconstructing activations from GPT-2 Small and Gemma 2 2B, and achieve comparable performance to state-of-the-art JumpReLU SAEs. However, an advantage of BatchTopK is that the average number of latents can be directly specified, rather than approximately tuned through a costly hyperparameter sweep. We provide code for training and evaluating BatchTopK SAEs at https://github.com/bartbussmann/BatchTopK

Key facts. Cited by 130 papers in the map (data through 14 Oct 2026). Placed in SAE features and their applications, one of 13 directions in Mechanistic interpretability. Builds on 5 of its 9 references that are in the map. First posted on arXiv 9 Dec 2024.

Builds on

5 of 9 references.

Closest ancestors · 4

Same direction, newest first

  1. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders19 Jul
  2. 2024Scaling and evaluating sparse autoencoders6 Jun
  3. 2024Improving Dictionary Learning with Gated Sparse Autoencoders24 Apr
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 1

  1. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 47; 2026: 83.

130 papers.

Sort:

By area:

  1. 2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 143
  2. 2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 80
  3. 2025Sparse Autoencoders Learn Monosemantic Features in Vision-Language Modelscited by 71
  4. 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometrycited by 61
  5. 2025Interpreting CLIP with Hierarchical Sparse Autoencoderscited by 60
  6. 2025From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuitcited by 49
  7. 2025Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuningcited by 34
  8. 2025Interpreting the linear structure of vision-language model embedding spacescited by 32
  9. 2026Do Sparse Autoencoders Capture Concept Manifolds?cited by 31
  10. 2025What's In My Human Feedback? Learning Interpretable Descriptions of Preference Datacited by 26
  11. 2026Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behaviorcited by 26
  12. 2025Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretabilitycited by 18
38 more
  1. 2025Feature Hedging: Correlated Features Break Narrow Sparse Autoencoderscited by 16
  2. 2025Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEscited by 14
  3. 2025Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoderscited by 14
  4. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?cited by 14
  5. 2025Analyze Feature Flow to Enhance Interpretation and Steering in Language Modelscited by 13
  6. 2025I Predict Therefore I Am: Is Next Token Prediction Enough to Learn Human-Interpretable Concepts from Data?cited by 13
  7. 2025Understanding sparse autoencoder scaling in the presence of feature manifoldscited by 11
  8. 2026SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Datacited by 11
  9. 2026Cross-Modal Redundancy and the Geometry of Vision-Language Embeddingscited by 10
  10. 2025AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Featurescited by 9
  11. 2025SAEdit: Token-level control for continuous image editing via Sparse AutoEncodercited by 9
  12. 2025Evaluating and Designing Sparse Autoencoders by Approximating Quasi-Orthogonalitycited by 8
  13. 2025OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Featurescited by 8
  14. 2025Low-Rank Adapting Models for Sparse Autoencoderscited by 7
  15. 2025Watch the Weights: Unsupervised monitoring and control of fine-tuned LLMscited by 7
  16. 2025Superposition as Lossy Compression: Measure with Sparse Autoencoders and Connect to Adversarial Vulnerabilitycited by 7
  17. 2025Route Experts by Sequence, not by Tokencited by 6
  18. 2025Interpretable and Steerable Concept Bottleneck Sparse Autoencoderscited by 6
  19. 2026Tracing Moral Foundations in Large Language Modelscited by 6
  20. 2026From Atoms to Trees: Building a Structured Feature Forest with Hierarchical Sparse Autoencoderscited by 6
  21. 2025Neurons Speak in Ranges: Breaking Free from Discrete Neuronal Attributioncited by 5
  22. 2025Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoderscited by 5
  23. 2025The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Featurescited by 5
  24. 2025Finding Manifolds With Bilinear Autoencoderscited by 5
  25. 2026Stable and Steerable Sparse Autoencoders with Weight Regularizationcited by 5
  26. 2026Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifoldscited by 5
  27. 2025Person-Centric Annotations of LAION-400M: Auditing Bias and Its Transfer to Modelscited by 4
  28. 2025Rank-1 LoRAs Encode Interpretable Reasoning Signalscited by 4
  29. 2026Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoderscited by 4
  30. 2026Are Sparse Autoencoder Benchmarks Reliable?cited by 4
  31. 2026Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretabilitycited by 4
  32. 2026ICA Lens: Interpreting Language Models Without Training Another Dictionarycited by 4
  33. 2025Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Modelscited by 3
  34. 2025Mechanistic Interpretability with Sparse Autoencoder Neural Operatorscited by 3
  35. 2025Binary Sparse Coding for Interpretabilitycited by 3
  36. 2025Sparse Autoencoders are Topic Modelscited by 3
  37. 2025A Geometric Unification of Concept Learning with Concept Conescited by 3
  38. 2026Towards Interpretable Hallucination Analysis and Mitigation in LVLMs via Contrastive Neuron Steeringcited by 3

50 most cited of 130 All 130 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025Distribution-Aware Feature Selection for SAEs
  2. 2025TopK Language Models
  3. 2025Sparse Autoencoders Trained on the Same Data Learn Different Features
  4. 2026Discovering Millions of Interpretable Features with Sparse Autoencoders
  5. 2025Revisiting End-To-End Sparse Autoencoder Training: A Short Finetune Is All You Need
  6. 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability
  7. 2026Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility
  8. 2025Train One Sparse Autoencoder Across Multiple Sparsity Budgets to Preserve Interpretability and Accuracy