mlsanityData to 7 Oct 2026

Programming Refusal with Conditional Activation Steering

arXiv 2409.05907 ↗First posted 6 Sep 2024cs.LG cs.AI cs.CLcited by 191

Not on the map. Why?

Cited by 48 in LLM safety and security, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

LLMs have shown remarkable capabilities, but precisely controlling their response behavior remains challenging. Existing activation steering methods alter LLM behavior indiscriminately, limiting their practical applicability in settings where selective responses are essential, such as content moderation or domain-specific assistants. In this paper, we propose Conditional Activation Steering (CAST), which analyzes LLM activation patterns during inference to selectively apply or withhold activation steering based on the input context. Our method is based on the observation that different categories of prompts activate distinct patterns in the model's hidden states. Using CAST, one can systematically control LLM behavior with rules like "if input is about hate speech or adult content, then refuse" or "if input is not about legal advice, then refuse." This allows for selective modification of responses to specific content while maintaining normal responses to other content, all without requiring weight optimization. We release an open-source implementation of our framework at github.com/IBM/activation-steering .

Key facts. Cited by 191 papers in the map (data through 7 Oct 2026). Builds on 70 of its 87 references that are in the map. First posted on arXiv 6 Sep 2024.

Builds on

70 of 87 references.

References · 70

  1. 2017Attention Is All You Needcited by 59,932
  2. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  3. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  4. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,203
  5. 2023Let's Verify Step by Stepcited by 3,598
  6. 2023Qwen Technical Reportcited by 3,273
  7. 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
  8. 2020Learning to summarize from human feedbackcited by 2,562
  9. 2022Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answeringcited by 1,881
  10. 2023WizardLM: Empowering large pre-trained language models to follow complex instructionscited by 1,116
  11. 2021A General Language Assistant as a Laboratory for Alignmentcited by 1,023
  12. 2023Whose Opinions Do Language Models Reflect?cited by 759
38 more
  1. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 679
  2. 2023Instruction Tuning with GPT-4cited by 676
  3. 2024OLMo: Accelerating the Science of Language Modelscited by 599
  4. 2023LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Datasetcited by 449
  5. 2023Language Models Represent Space and Timecited by 306
  6. 2023MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Modelscited by 255
  7. 2024MM1: Methods, Analysis & Insights from Multimodal LLM Pre-trainingcited by 230
  8. 2024SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusalcited by 224
  9. 2024OR-Bench: An Over-Refusal Benchmark for Large Language Modelscited by 222
  10. 2020ReClor: A Reading Comprehension Dataset Requiring Logical Reasoningcited by 211
  11. 2024Foundational Challenges in Assuring Alignment and Safety of Large Language Modelscited by 193
  12. 2024ReFT: Representation Finetuning for Language Modelscited by 186
  13. 2023Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMscited by 176
  14. 2024A Roadmap to Pluralistic Alignmentcited by 175
  15. 2023Platypus: Quick, Cheap, and Powerful Refinement of LLMscited by 150
  16. 2024The Remarkable Robustness of LLMs: Stages of Inference?cited by 138
  17. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 130
  18. 2024Safe LoRA: the Silver Lining of Reducing Safety Risks when Fine-tuning Large Language Modelscited by 122
  19. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 112
  20. 2024Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaborationcited by 111
  21. 2024The Art of Saying No: Contextual Noncompliance in Language Modelscited by 100
  22. 2024Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categoriescited by 74
  23. 2024Transformers represent belief state geometry in their residual streamcited by 73
  24. 2024Roleplay-doh: Enabling Domain-Experts to Create LLM-simulated Patients via Eliciting and Adhering to Principlescited by 72
  25. 2023Word Embeddings Are Steers for Language Modelscited by 70
  26. 2023Improving Activation Steering in Language Models with Mean-Centringcited by 64
  27. 2024Aligning Large Language Models with Representation Editing: A Control Perspectivecited by 62
  28. 2024(A)I Am Not a Lawyer, But...: Engaging Legal Experts towards Responsible LLM Policies for Legal Advicecited by 61
  29. 2023Concept Algebra for (Score-Based) Text-Controlled Generative Modelscited by 59
  30. 2022Can Large Language Models Truly Understand Prompts? A Case Study with Negated Promptscited by 56
  31. 2024Style Vectors for Steering Generative Large Language Modelcited by 56
  32. 2024LoFiT: Localized Fine-tuning on LLM Representationscited by 56
  33. 2022Transcending Scaling Laws with 0.1% Extra Computecited by 55
  34. 2024Spectral Editing of Activations for Large Language Model Alignmentcited by 54
  35. 2023Controlled Text Generation via Language Model Arithmeticcited by 51
  36. 2024Steering Without Side Effects: Improving Post-Deployment Control of Language Modelscited by 44
  37. 2023Specific versus General Principles for Constitutional AIcited by 41
  38. 2024Who's asking? User personas and the mechanics of latent misalignmentcited by 40

50 most cited of 70 All 70 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 7; 2025: 57; 2026: 127.

191 papers.

Sort:

By area:

  1. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 89
  2. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 86
  3. 2025AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraintcited by 45
  4. 2025Angular Steering: Behavior Control via Rotation in Activation Spacecited by 45
  5. 2025A Unified Understanding and Evaluation of Steering Methodscited by 40
  6. 2025Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safetycited by 38
  7. 2025AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defendercited by 37
  8. 2025No Answer Needed: Predicting LLM Answer Accuracy from Question-Only Linear Probescited by 35
  9. 2024Evaluating the Prompt Steerability of Large Language Modelscited by 26
  10. 2025To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Modelscited by 26
  11. 2024Locking Down the Finetuned LLMs Safetycited by 22
  12. 2025Representation Engineering for Large-Language Models: Survey and Research Challengescited by 18
38 more
  1. 2025Spotlight Your Instructions: Instruction-following with Dynamic Attention Steeringcited by 18
  2. 2025Mitigating Content Effects on Reasoning in Language Models through Fine-Grained Activation Steeringcited by 18
  3. 2026Spherical Steering: Geometry-Aware Activation Rotation for Language Modelscited by 17
  4. 2026A Framework for Formalizing LLM Agent Securitycited by 16
  5. 2025Differentially Private Steering for Large Language Model Alignmentcited by 15
  6. 2025COSMIC: Generalized Refusal Direction Identification in LLM Activationscited by 15
  7. 2026ODESteer: A Unified ODE-Based Steering Framework for LLM Alignmentcited by 15
  8. 2025Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Controlcited by 14
  9. 2025Anthropomimetic Uncertainty: What Verbalized Uncertainty in Language Models is Missingcited by 14
  10. 2025MMPB: It's Time for Multi-Modal Personalizationcited by 14
  11. 2025Learning to Steer: Input-dependent Steering for Multimodal LLMscited by 13
  12. 2025EasySteer: A Unified Framework for High-Performance and Extensible LLM Steeringcited by 13
  13. 2024Refusal Tokens: A Simple Way to Calibrate Refusals in Large Language Modelscited by 12
  14. 2025Activation Steering with a Feedback Controllercited by 11
  15. 2026Understanding LoRA as Knowledge Memory: An Empirical Analysiscited by 11
  16. 2025Jailbreak Attack Initializations as Extractors of Compliance Directionscited by 10
  17. 2025RepIt: Steering Language Models with Concept-Specific Refusal Vectorscited by 10
  18. 2025MONICA: Real-Time Monitoring and Calibration of Chain-of-Thought Sycophancy in Large Reasoning Modelscited by 10
  19. 2025MLLMEraser: Achieving Test-Time Unlearning in Multimodal Large Language Models through Activation Steeringcited by 9
  20. 2025Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Modelscited by 8
  21. 2025Guiding Giants: Lightweight Controllers for Weighted Activation Steering in LLMscited by 8
  22. 2025KV Cache Steering for Controlling Frozen LLMscited by 8
  23. 2025Guardians and Offenders: A Survey on Harmful Content Generation and Safety Mitigation of LLMcited by 8
  24. 2025Speech-Audio Compositional Attacks on Multimodal LLMs and Their Mitigation with SALMONN-Guardcited by 8
  25. 2026Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Modelscited by 8
  26. 2026Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Controlcited by 8
  27. 2025Unsupervised Concept Vector Extraction for Bias Control in LLMscited by 7
  28. 2025Effectively Steer LLM To Follow Preference via Building Confident Directionscited by 7
  29. 2025Steering When Necessary: Flexible Steering Large Language Models with Backtrackingcited by 7
  30. 2025Steering Protein Language Modelscited by 7
  31. 2025Activation Steering for Masked Diffusion Language Modelscited by 7
  32. 2026What Drives Representation Steering? A Mechanistic Case Study on Steering Refusalcited by 7
  33. 2026Steered LLM Activations are Non-Surjectivecited by 7
  34. 2025Safety Subspaces are Not Linearly Distinct: A Fine-Tuning Case Studycited by 6
  35. 2025When Do LLMs Admit Their Mistakes? Understanding The Role Of Model Belief In Retractioncited by 6
  36. 2025What Can We Actually Steer? A Multi-Behavior Study of Activation Controlcited by 6
  37. 2026ASA: Backbone-Training-Free Representation Engineering for Tool-Calling Agentscited by 6
  38. 2026Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMscited by 6

50 most cited of 191 All 191 (JSON)