mlsanityData to 7 Oct 2026

Representation Surgery: Theory and Practice of Affine Steering

arXiv 2402.09631 ↗First posted 15 Feb 2024cs.LG cs.CL cs.CYcited by 47

Not on the map. Why?

Abstract (arXiv)

Language models often exhibit undesirable behavior, e.g., generating toxic or gender-biased text. In the case of neural language models, an encoding of the undesirable behavior is often present in the model's representations. Thus, one natural (and common) approach to prevent the model from exhibiting undesirable behavior is to steer the model's representations in a manner that reduces the probability of it generating undesirable text. This paper investigates the formal and empirical properties of steering functions, i.e., transformation of the neural language model's representations that alter its behavior. First, we derive two optimal, in the least-squares sense, affine steering functions under different constraints. Our theory provides justification for existing approaches and offers a novel, improved steering approach. Second, we offer a series of experiments that demonstrate the empirical effectiveness of the methods in mitigating bias and reducing toxic generation.

Key facts. Cited by 47 papers in the map (data through 7 Oct 2026). Builds on 28 of its 43 references that are in the map. First posted on arXiv 15 Feb 2024.

Builds on

28 of 43 references.

References · 28

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 13,958
  3. 2016Pointer Sentinel Mixture Modelscited by 2,992
  4. 2022Locating and Editing Factual Associations in GPTcited by 2,609
  5. 2021Convex optimizationcited by 2,479
  6. 2020Transformer Feed-Forward Layers Are Key-Value Memoriescited by 1,307
  7. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,304
  8. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  9. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,235
  10. 2019Universal Adversarial Triggers for Attacking and Analyzing NLPcited by 709
  11. 2019The Woman Worked as a Babysitter: On Biases in Language Generationcited by 540
  12. 2020Null It Out: Guarding Protected Attributes by Iterative Nullspace Projectioncited by 442
16 more
  1. 2021DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Expertscited by 434
  2. 2019Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Settingcited by 346
  3. 2019Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Themcited by 268
  4. 2023LEACE: Perfect linear concept erasure in closed formcited by 254
  5. 2024Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Modelscited by 225
  6. 2022Extracting Latent Steering Vectors from Pretrained Language Modelscited by 217
  7. 2016Demographic Dialectal Variation in Social Media: A Case Study of African-American Englishcited by 194
  8. 2018Learning Gender-Neutral Word Embeddingscited by 169
  9. 2018Adversarial Removal of Demographic Attributes from Text Datacited by 166
  10. 2020CausaLM: Causal Model Explanation Through Counterfactual Language Modelscited by 117
  11. 2022Fair and Optimal Classification via Post-Processingcited by 50
  12. 2022Kernelized Concept Erasurecited by 47
  13. 2022Unified Detoxifying and Debiasing in Language Generation via Inference-time Adaptive Optimizationcited by 32
  14. 2023A Geometric Notion of Causal Probingcited by 23
  15. 2023Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Modelscited by 23
  16. 2021Domain-Adaptive Pretraining Methods for Dialogue Understandingcited by 11

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 7; 2025: 17; 2026: 23.

47 papers.

Sort:

By area:

  1. 2024ReFT: Representation Finetuning for Language Modelscited by 186
  2. 2024A Primer on the Inner Workings of Transformer-based Language Modelscited by 83
  3. 2024Spectral Editing of Activations for Large Language Model Alignmentcited by 54
  4. 2024LLaVA Steering: Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steeringcited by 50
  5. 2025Angular Steering: Behavior Control via Rotation in Activation Spacecited by 45
  6. 2025A Unified Understanding and Evaluation of Steering Methodscited by 40
  7. 2025Understanding (Un)Reliability of Steering Vectors in Language Modelscited by 39
  8. 2024The Quest for the Right Mediator: Surveying Mechanistic Interpretability Through the Lens of Causal Mediation Analysiscited by 24
  9. 2024Controllable Context Sensitivity and the Knob Behind Itcited by 22
  10. 2026ODESteer: A Unified ODE-Based Steering Framework for LLM Alignmentcited by 15
  11. 2024Gumbel Counterfactual Generation From Language Modelscited by 13
  12. 2025SAKE: Steering Activations for Knowledge Editingcited by 12
35 more
  1. 2025Emergence of Linear Truth Encodings in Language Modelscited by 12
  2. 2025Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safetycited by 10
  3. 2026From Directions to Regions: Decomposing Activations in Language Models via Local Geometrycited by 10
  4. 2026The Information Geometry of Softmax: Probing and Steeringcited by 9
  5. 2026Can LLMs Introspect? A Reality Checkcited by 8
  6. 2025Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineeringcited by 7
  7. 2025ReCoVeR the Target Language: Language Steering without Sacrificing Task Performancecited by 5
  8. 2025DISCO: Disentangled Communication Steering for Large Language Modelscited by 5
  9. 2026Curveball Steering: The Right Direction To Steer Isn't Always Linearcited by 5
  10. 2025Toward a Flexible Framework for Linear Representation Hypothesis Using Maximum Likelihood Estimationcited by 3
  11. 2025Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigationcited by 3
  12. 2026Concept Heterogeneity-aware Representation Steeringcited by 3
  13. 2026Where to Steer: Input-Dependent Layer Selection for Steering Improves LLM Alignmentcited by 3
  14. 2026FishBack: Pullback Fisher Geometry for Optimal Activation Steering in Transformerscited by 3
  15. 2025Distributional Surgery for Language Model Activationscited by 2
  16. 2025Abstract Counterfactuals for Language Model Agentscited by 2
  17. 2025Nonlinear Concept Erasure: a Density Matching Approachcited by 2
  18. 2025On Optimal Steering to Achieve Exact Fairnesscited by 2
  19. 2026Time Travel Engine: A Shared Latent Chronological Manifold Enables Historical Navigation in Large Language Modelscited by 2
  20. 2026Don't Lose Focus: Activation Steering via Key-Orthogonal Projectionscited by 2
  21. 2026Conceptors for Semantic Steeringcited by 1
  22. 2026Conditional Optimal Bridge for Riemannian Activation Steeringcited by 1
  23. 2025Group Fairness Meets the Black Box: Enabling Fair Algorithms on Closed LLMs via Post-Processingcited by 0
  24. 2025Sy-FAR: Symmetry-based Fair Adversarial Robustnesscited by 0
  25. 2026BarrierSteer: LLM Safety via Learning Barrier Steeringcited by 0
  26. 2026ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Modelscited by 0
  27. 2026MANCE: Manifold Aware Concept Erasurecited by 0
  28. 2026Safety Cost of Steering Vectors Is Separable and Reduciblecited by 0
  29. 2026Predictive Memory Localization: Forecasting Selective Intervention Paths from Internal Signalscited by 0
  30. 2026GEOSTEER: Geodesic Optimization for Activation Steering in Large Language Modelscited by 0
  31. 2026MUtE: A Dual Framework for Concept Erasure and Counterfactual Interventionscited by 0
  32. 2026Replay-Gated Neural Execution: Decoupling Persistent Behavioral Specifications from Neural Realizations in Frozen Language Modelscited by 0
  33. 2026Shared Weights, Selected Computations: How Looped Transformers Route What Each Loop Doescited by 0
  34. 2026Capturing In-Context Learning Dynamics with Task Operatorscited by 0
  35. 2026Kernelized Activation Steeringcited by 0