Data to 7 Oct 2026

Sparse Autoencoders Find Highly Interpretable Features in Language Models

arXiv 2309.08600 ↗First posted 15 Sep 2023cs.LG cs.CLcited by 1,359

Not on the map. Why?

Cited by 72 in LLM safety and security, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

One of the roadblocks to a better understanding of neural networks' internals is textitpolysemanticity, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is textitsuperposition, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Moreover, we show that with our learned set of features, we can pinpoint the features that are causally responsible for counterfactual behaviour on the indirect object identification task citepwang2022interpretability to a finer degree than previous decompositions. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability.

Key facts. Cited by 1,359 papers in the map (data through 7 Oct 2026). Builds on 19 of its 38 references that are in the map. First posted on arXiv 15 Sep 2023.

Builds on

19 of 38 references.

References · 19

2020The Pile: An 800GB Dataset of Diverse Text for Language Modelingcited by 2,346
2018The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networkscited by 2,250
2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,734
2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,130
2022Toy Models of Superpositioncited by 928
2022LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scalecited by 881
2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 700
2021Causal Abstractions of Neural Networkscited by 401
2023LEACE: Perfect linear concept erasure in closed formcited by 254
2022The Alignment Problem from a Deep Learning Perspectivecited by 246
2023An Overview of Catastrophic AI Riskscited by 214
2019Adaptively Sparse Transformerscited by 154

7 more

2021Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factorscited by 106
2022Linear Adversarial Concept Erasurecited by 98
2017SPINE: SParse Interpretable Neural Embeddingscited by 55
2015Sparse Overcomplete Word Vector Representationscited by 43
2018Accelerating Convolutional Neural Networks via Activation Map Compressioncited by 34
2021Is Sparse Attention more Interpretable?cited by 31
2019Analysis of the Optimization Landscapes for Overcomplete Representation Learningcited by 7

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2023: 14; 2024: 177; 2025: 483; 2026: 685.

1,359 papers.

Sort:

By area:

2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 673
2024KAN: Kolmogorov-Arnold Networkscited by 634
2024Scaling and evaluating sparse autoencoderscited by 530
2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 415
2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 412
2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 381
2023Towards Best Practices of Activation Patching in Language Models: Metrics and Methodscited by 328
2023Language Models Represent Space and Timecited by 306
2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 283
2024Not All Language Model Features Are One-Dimensionally Linearcited by 201
2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 199

38 more

2024How to use and interpret activation patchingcited by 191
2024Transcoders Find Interpretable LLM Feature Circuitscited by 186
2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 171
2024Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMscited by 158
2024The Geometry of Categorical and Hierarchical Concepts in Large Language Modelscited by 151
2024Improving Instruction-Following in Language Models through Activation Steeringcited by 149
2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 148
2025Learning Multi-Level Features with Matryoshka Sparse Autoencoderscited by 136
2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 133
2025A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyondcited by 131
2024BatchTopK Sparse Autoencoderscited by 126
2024Black-Box Access is Insufficient for Rigorous AI Auditscited by 123
2024Wav-KAN: Wavelet Kolmogorov-Arnold Networkscited by 118
2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretabilitycited by 118
2024KAN 2.0: Kolmogorov-Arnold Networks Meet Sciencecited by 116
2024Towards Interpreting Visual Information Processing in Vision-Language Modelscited by 115
2024A Survey on Uncertainty Quantification of Large Language Models: Taxonomy, Open Research Challenges, and Future Directionscited by 115
2024Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Modelscited by 114
2024A Practical Review of Mechanistic Interpretability for Transformer-Based Language Modelscited by 113
2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 109
2024Automatically Interpreting Millions of Features in Large Language Modelscited by 106
2025Agentic Large Language Models, a surveycited by 105
2025The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independencecited by 104
2024Interpreting and Editing Vision-Language Representations to Mitigate Hallucinationscited by 103
2024Universal Neurons in GPT2 Language Modelscited by 100
2024AtP*: An efficient and scalable method for localizing LLM behaviour to componentscited by 100
2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
2024Steering Language Model Refusal with Sparse Autoencoderscited by 89
2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 86
2024A Primer on the Inner Workings of Transformer-based Language Modelscited by 83
2024Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Controlcited by 83
2024Truth is Universal: Robust Detection of Lies in LLMscited by 78
2023Successor Heads: Recurring, Interpretable Attention Heads In The Wildcited by 77
2024RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representationscited by 77
2025Persona Features Control Emergent Misalignmentcited by 76
2025SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoderscited by 75
2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 74
2025A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Modelscited by 73

50 most cited of 1,359 All 1,359 (JSON)