mlsanityData to 14 Oct 2026

SAEs Are Good for Steering -- If You Select the Right Features

arXiv 2505.20063 ↗First posted 26 May 2025cs.LG cs.AI cs.CLcited by 78

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

Abstract (arXiv)

Sparse Autoencoders (SAEs) have been proposed as an unsupervised approach to learn a decomposition of a model's latent space. This enables useful applications such as steering - influencing the output of a model towards a desired concept - without requiring labeled data. Current methods identify SAE features to steer by analyzing the input tokens that activate them. However, recent work has highlighted that activations alone do not fully describe the effect of a feature on the model's output. In this work, we draw a distinction between two types of features: input features, which mainly capture patterns in the model's input, and output features, which have a human-understandable effect on the model's output. We propose input and output scores to characterize and locate these types of features, and show that high values for both scores rarely co-occur in the same features. These findings have practical implications: after filtering out features with low output scores, we obtain 2-3x improvements when steering with SAEs, making them competitive with supervised methods.

Key facts. Cited by 78 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 36 of its 66 references that are in the map. First posted on arXiv 26 May 2025.

Builds on

36 of 66 references.

Closest ancestors · 4

Same direction, newest first

  1. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders28 Jan
  2. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders27 Oct
  3. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders19 Jul
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 32

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  3. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  4. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  5. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,786
  6. 2020Transformer Feed-Forward Layers Are Key-Value Memoriescited by 1,336
  7. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  8. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  9. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  10. 2019BERT Rediscovers the Classical NLP Pipelinecited by 1,059
  11. 2023Steering Language Models With Activation Engineeringcited by 922
  12. 2023Dissecting Recall of Factual Associations in Auto-Regressive Language Modelscited by 528
20 more
  1. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  2. 2015Autoencoding beyond pixels using a learned similarity metriccited by 407
  3. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  4. 2019Linguistic Knowledge and Transferability of Contextual Representationscited by 383
  5. 2023In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steeringcited by 245
  6. 2024ReFT: Representation Finetuning for Language Modelscited by 197
  7. 2017What do Neural Machine Translation Models Learn about Morphology?cited by 166
  8. 2024The Remarkable Robustness of LLMs: Stages of Inference?cited by 146
  9. 2019On Identifiability in Transformerscited by 116
  10. 2018Identifying and Controlling Important Neurons in Neural Machine Translationcited by 111
  11. 2024Universal Neurons in GPT2 Language Modelscited by 102
  12. 2025Enhancing Automated Interpretability with Output-Centric Feature Descriptionscited by 41
  13. 2024Extending Activation Steering to Broad Skills and Multiple Behaviourscited by 40
  14. 2025FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steeringcited by 40
  15. 2023ReFACT: Updating Text-to-Image Models by Editing the Text Encodercited by 32
  16. 2024Backward Lens: Projecting Language Model Gradients into the Vocabulary Spacecited by 32
  17. 2024Style-Specific Neurons for Steering LLMs in Text Style Transfercited by 24
  18. 2024Constructing Benchmarks and Interventions for Combating Hallucinations in LLMscited by 19
  19. 2018Language Modeling Teaches You More Syntax than Translation Does: Lessons Learned Through Auxiliary Task Analysiscited by 17
  20. 2024Using Shapley interactions to understand how models use structurecited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 18; 2026: 60.

78 papers.

Sort:

By area:

  1. 2025Persona Features Control Emergent Misalignmentcited by 86
  2. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  3. 2025The Rogue Scalpel: Activation Steering Compromises LLM Safetycited by 21
  4. 2026Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Modelscited by 21
  5. 2025Position: Use Sparse Autoencoders to Discover Unknownscited by 19
  6. 2025Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretabilitycited by 18
  7. 2025CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Featurescited by 12
  8. 2025CRISP: Persistent Concept Unlearning via Sparse Autoencoderscited by 11
  9. 2026The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patchingcited by 11
  10. 2025Enhancing LLM Steering through Sparse Autoencoder-Based Vector Refinementcited by 10
  11. 2025Does higher interpretability imply better utility? A Pairwise Analysis on Sparse Autoencoderscited by 10
  12. 2025SAEdit: Token-level control for continuous image editing via Sparse AutoEncodercited by 9
38 more
  1. 2026From Weights to Activations: Is Steering the Next Frontier of Adaptation?cited by 8
  2. 2025From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?cited by 7
  3. 2026DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoderscited by 7
  4. 2025Interpretable and Steerable Concept Bottleneck Sparse Autoencoderscited by 6
  5. 2026Post-Hoc Reasoning in Chain of Thought: Decoding and Steering Pre-Committed Answerscited by 6
  6. 2026Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lenscited by 6
  7. 2026Steering at the Source: Style Modulation Heads for Robust Persona Controlcited by 5
  8. 2025Evaluating Sparse Autoencoders for Monosemantic Representationcited by 4
  9. 2025AlignSAE: Concept-Aligned Sparse Autoencoderscited by 4
  10. 2026The Confidence Manifold: Geometric Structure of Correctness Representations in Language Modelscited by 4
  11. 2026Interpreting and Steering State-Space Models via Activation Subspace Bottleneckscited by 3
  12. 2026Steering Without Breaking: Mechanistically Informed Interventions for Discrete Diffusion Language Modelscited by 3
  13. 2026Steering LLMs? Actually, Sparse Autoencoders can outperform simple baselinescited by 3
  14. 2025CE-Bench: Towards a Reliable Contrastive Evaluation Benchmark of Interpretability of Sparse Autoencoderscited by 2
  15. 2025Can SAEs reveal and mitigate racial biases of LLMs in healthcare?cited by 2
  16. 2026Geometric Routing Enables Causal Expert Control in Mixture of Expertscited by 2
  17. 2026Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligencecited by 2
  18. 2025Calibrating Lightweight Sparse Autoencoder Feature Steeringcited by 1
  19. 2025Measuring Mechanistic Independence: Can Bias Be Removed Without Erasing Demographics?cited by 1
  20. 2026DSPA: Dynamic SAE Steering for Data-Efficient Preference Alignmentcited by 1
  21. 2026Auditing Asset-Specific Preferences in Financial Large Language Models: Evidence from Bitcoin Representations and Portfolio Allocationcited by 1
  22. 2026SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behaviorcited by 1
  23. 2026Do Sparse Autoencoders Learn Meaningful Concept Hierarchies?cited by 1
  24. 2026Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effectscited by 1
  25. 2026Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effectscited by 1
  26. 2026CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuitscited by 1
  27. 2026REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Featurescited by 1
  28. 2025When Models Refuse: Political Steerability and Feature Richness as Measures of Ideological Depthcited by 0
  29. 2026Mechanistic Indicators of Steering Effectiveness in Large Language Modelscited by 0
  30. 2026Sparse Visual Thought Circuits in Vision-Language Modelscited by 0
  31. 2026A Universal Vibe? Finding and Controlling Language-Agnostic Informal Register with SAEscited by 0
  32. 2026Multimodal QUD: Inquisitive Questions from Scientific Figurescited by 0
  33. 2026MoRFI: Monotonic Sparse Autoencoder Feature Identificationcited by 0
  34. 2026Steering grids for sparse-autoencoder features: when a top-context label names an activation regime rather than a causal axiscited by 0
  35. 2026Domain Restriction via Multi SAE Layer Transitionscited by 0
  36. 2026Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Modelscited by 0
  37. 2026Steered Generation via Gradient-Based Optimization on Sparse Query Featurescited by 0
  38. 2026Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Modelscited by 0

50 most cited of 78 All 78 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2024Can sparse autoencoders be used to decompose and interpret steering vectors?
  2. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Features
  3. 2026Enhancing SAE-based Steering via Neighbor Integrated Feature Selection
  4. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
  5. 2025Sparse Autoencoders Trained on the Same Data Learn Different Features
  6. 2024Features that Make a Difference: Leveraging Gradients for Improved Dictionary Learning
  7. 2024Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks
  8. 2025SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Models