mlsanityData to 14 Oct 2026

AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders

arXiv 2501.17148 ↗First posted 28 Jan 2025cs.CL cs.AI cs.LGcited by 218

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

Abstract (arXiv)

Fine-grained steering of language model outputs is essential for safety and reliability. Prompting and finetuning are widely used to achieve these goals, but interpretability researchers have proposed a variety of representation-based techniques as well, including sparse autoencoders (SAEs), linear artificial tomography, supervised steering vectors, linear probes, and representation finetuning. At present, there is no benchmark for making direct comparisons between these proposals. Therefore, we introduce AxBench, a large-scale benchmark for steering and concept detection, and report experiments on Gemma-2-2B and 9B. For steering, we find that prompting outperforms all existing methods, followed by finetuning. For concept detection, representation-based methods such as difference-in-means, perform the best. On both evaluations, SAEs are not competitive. We introduce a novel weakly-supervised representational method (Rank-1 Representation Finetuning; ReFT-r1), which is competitive on both tasks while providing the interpretability advantages that prompting lacks. Along with AxBench, we train and publicly release SAE-scale feature dictionaries for ReFT-r1 and DiffMean.

Key facts. Cited by 218 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 42 of its 63 references that are in the map. First posted on arXiv 28 Jan 2025.

Builds on

42 of 63 references.

Closest ancestors · 5

Same direction, newest first

  1. 2024Can sparse autoencoders be used to decompose and interpret steering vectors?13 Nov
  2. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Features4 Nov
  3. 2024Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoders9 Oct
  4. 2024Scaling and evaluating sparse autoencoders6 Jun
  5. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 37

  1. 2013Distributed Representations of Words and Phrases and their Compositionalitycited by 3,474
  2. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  3. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  4. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  5. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,312
  6. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  7. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  8. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  9. 2022Toy Models of Superpositioncited by 971
  10. 2023Steering Language Models With Activation Engineeringcited by 922
  11. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
  12. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
25 more
  1. 2018What you can cram into a single vector: Probing sentence embeddings for linguistic propertiescited by 457
  2. 2021Causal Abstractions of Neural Networkscited by 422
  3. 2015Autoencoding beyond pixels using a learned similarity metriccited by 407
  4. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  5. 2024Improving Alignment and Robustness with Circuit Breakerscited by 322
  6. 2023Emergent Linear Representations in World Models of Self-Supervised Sequence Modelscited by 319
  7. 2023In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steeringcited by 245
  8. 2022Extracting Latent Steering Vectors from Pretrained Language Modelscited by 225
  9. 2023Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretabilitycited by 205
  10. 2024ReFT: Representation Finetuning for Language Modelscited by 197
  11. 2023Localizing Model Behavior with Path Patchingcited by 195
  12. 2017What do Neural Machine Translation Models Learn about Morphology?cited by 166
  13. 2021Inducing Causal Structure for Interpretable Neural Networkscited by 110
  14. 2019Implicit Semantic Data Augmentation for Deep Networkscited by 106
  15. 2018Under the Hood: Using Diagnostic Classifiers to Investigate and Improve how Language Models Track Agreement Informationcited by 97
  16. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 95
  17. 2024Defending Against Unforeseen Failure Modes with Latent Adversarial Trainingcited by 88
  18. 2016Deep Feature Interpolation for Image Content Changescited by 80
  19. 2019AllenNLP Interpret: A Framework for Explaining Predictions of NLP Modelscited by 74
  20. 2024pyvene: A Library for Understanding and Improving PyTorch Models via Interventionscited by 54
  21. 2024Extending Activation Steering to Broad Skills and Multiple Behaviourscited by 40
  22. 2024Towards Reliable Evaluation of Behavior Steering Interventions in LLMscited by 36
  23. 2024Towards Unifying Interpretability and Control: Evaluation via Interventioncited by 32
  24. 2023A Geometric Notion of Causal Probingcited by 25
  25. 2023Understanding and Controlling a Maze-Solving Policy Networkcited by 22

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 79; 2026: 139.

218 papers.

Sort:

By area:

  1. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 115
  2. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
  3. 2025A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Modelscited by 76
  4. 2025SEAL: Steerable Reasoning Calibration of Large Language Models for Freecited by 68
  5. 2025Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometrycited by 61
  6. 2025MIB: A Mechanistic Interpretability Benchmarkcited by 56
  7. 2025Detecting High-Stakes Interactions with Activation Probescited by 53
  8. 2025Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuningcited by 39
  9. 2025Sycophancy Is Not One Thing: Causal Separation of Sycophantic Behaviors in LLMscited by 37
  10. 2025Steering off Course: Reliability Challenges in Steering Language Modelscited by 34
  11. 2026Do Sparse Autoencoders Capture Concept Manifolds?cited by 31
  12. 2025SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Modelscited by 29
38 more
  1. 2025To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Modelscited by 29
  2. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  3. 2026Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behaviorcited by 26
  4. 2025Beyond Linear Steering: Unified Multi-Attribute Control for Language Modelscited by 25
  5. 2026Language Model Circuits Are Sparse in the Neuron Basiscited by 25
  6. 2025Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atomscited by 24
  7. 2025HyperSteer: Activation Steering at Scale with Hypernetworkscited by 23
  8. 2025Sparse Autoencoder Features for Classifications and Transferabilitycited by 21
  9. 2025Improved Representation Steering for Language Modelscited by 21
  10. 2025The Rogue Scalpel: Activation Steering Compromises LLM Safetycited by 21
  11. 2025Position: Use Sparse Autoencoders to Discover Unknownscited by 19
  12. 2025Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretabilitycited by 18
  13. 2025Feature Hedging: Correlated Features Break Narrow Sparse Autoencoderscited by 16
  14. 2025Aligned but Blind: Alignment Increases Implicit Bias by Reducing Awareness of Racecited by 15
  15. 2025Evaluation Faking: Unveiling Observer Effects in Safety Evaluation of Frontier AI Systemscited by 14
  16. 2025Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoderscited by 14
  17. 2026Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?cited by 14
  18. 2025Discovering and Steering Interpretable Concepts in Large Generative Music Modelscited by 13
  19. 2026Shaping capabilities with token-level data filteringcited by 13
  20. 2026Steering Vector Fields for Context-Aware Inference-Time Control in Large Language Modelscited by 13
  21. 2025RepIt: Steering Language Models with Concept-Specific Refusal Vectorscited by 12
  22. 2026CLaS-Bench: A Cross-Lingual Alignment and Steering Benchmarkcited by 12
  23. 2026From Directions to Regions: Decomposing Activations in Language Models via Local Geometrycited by 12
  24. 2025Explainability in Practice: A Survey of Explainable NLP Across Various Domainscited by 11
  25. 2025Precise In-Parameter Concept Erasure in Large Language Modelscited by 11
  26. 2025BILLY: Steering Large Language Models via Merging Persona Vectors for Creative Generationcited by 11
  27. 2026The Information Geometry of Softmax: Probing and Steeringcited by 11
  28. 2025SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMscited by 10
  29. 2025Measuring and Guiding Monosemanticitycited by 10
  30. 2025Enhancing LLM Steering through Sparse Autoencoder-Based Vector Refinementcited by 10
  31. 2025Does higher interpretability imply better utility? A Pairwise Analysis on Sparse Autoencoderscited by 10
  32. 2026Analysing the Safety Pitfalls of Steering Vectorscited by 10
  33. 2025Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Modelscited by 9
  34. 2025Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoderscited by 9
  35. 2025Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steeringcited by 9
  36. 2025AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Featurescited by 9
  37. 2026Why Steering Works: Toward a Unified View of Language Model Parameter Dynamicscited by 9
  38. 2025Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoderscited by 8

50 most cited of 218 All 218 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025Interpretable LLM Guardrails via Sparse Representation Steering
  2. 2026Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects
  3. 2025Steering Large Language Model Activations in Sparse Spaces
  4. 2025CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features
  5. 2026SwordBench: Evaluating Orthogonality of Steering Image Representations
  6. 2025Feature-Guided SAE Steering for Refusal-Rate Control using Contrasting Prompts
  7. 2026Enhancing SAE-based Steering via Neighbor Integrated Feature Selection
  8. 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability