mlsanityData to 14 Oct 2026

Activation steering methods

239 papers5 this weekMedium confidence in this grouping: mostly one topic, with blurry edgesMedium confidence in this grouping: mostly one topic, with blurry edges · About

Keywords: steering, activation, activation steering, language, vectors

Key facts. Activation steering methods: 239 papers on arXiv since 2022 (data through 14 Oct 2026). 159 of its 239 mapped papers (67%) cite “Steering Llama 2 via Contrastive Activation Addition” (2023, arXiv 2312.06681). 28 new papers 24 Sep–14 Oct.

What’s new

24 Sep – 14 Oct 2026

28 new papers, 24 Sep – 14 Oct; 5 this week.

  1. This week · 5
  2. Causal Syntax Circuits in CodeLMs: Budget and Probe Decide What Looks Localised also fits: Knowledge and representations inside LLMs
  3. Genesis Steering: Better Activation Steering Through Training Data
  4. Perturbation, Not Provenance: Why Activation Steering Does Not Make a Robust Text Watermark
  5. Beyond Bias Scores: Answerability Confounds in Activation Steering with Gradient-Constrained Direction Design (GCD) also fits: Probing deception, bias and beliefs
  6. The Sharp Directions Are Against You: Curvature Analysis of Activation Steering
  7. Earlier · 23
  8. What Survives When Activation Steering Is Used as Memory
  9. Steerspeech: Activation Steering For Emotion Control In Generated Speech
  10. Steering Follows Geometry, Not Labels: Emotion Directions in a Full-Duplex Speech Model
  11. Does Steering Break Your Model? A Multi-Dimensional Evaluation Suite for LLM Steering Methods
  12. Loud and Clear: Dynamic Activation Steering for Improving Speech Intelligibility in Noisy Environments
All 28
  1. Steering by Influence: Curvature Aware Data Weighting for Activation Steering
  2. Tracing a Sparse Emotion-Control Circuit in LLM-Based Text-to-Speech
  3. One Token Can Be Enough: Bridging Prompting and Activation Steering with Prefix Steering
  4. Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recovery also fits: Sparse autoencoder training and evaluation
  5. Steering Speech-Language Models: Training-Free Task Specialization via Contrastive Activation Addition
  6. First-Order Steering: Translating Weight Adaptation into Activation Steering
  7. On the Steering Dimensionality of Refusal in Language Models
  8. Predicting Steering Vectors and Adapter Weights for Few-Shot Author-Style Transfer
  9. Hesitation Has a Geometry: Entropy-Trained Hyperbolic Probes for Sparse Activation Steering also fits: Steering and interpreting LLM reasoning
  10. Kernelized Activation Steering
  11. Component and Dimension Sparsity in Transformer Refusal Mechanisms also fits: Steering and probing for LLM safety
  12. Steering Fields: Adaptive Vector Fields for Safe Image Generation and Beyond also fits: Interpreting and steering diffusion models
  13. Contrastive Covariance Subspaces for Activation Steering
  14. Persona Dosing: Calibrated Activation Steering for Graded Trait Control
  15. Controlling Speaking Rate in Autoregressive TTS via Activation Steering
  16. LocUS: Head Selection and Subspace Projection for Targeted Activation Steering
  17. The Linear Representation Hypothesis for Vision-Language-Action Models also fits: Knowledge and representations inside LLMs
  18. Adaptive Multi-Value Control in LLMs via Causal Activation Steering

Foundations

Most cited by papers here

  1. Steering Llama 2 via Contrastive Activation Addition2023cited by 159 here
  2. Steering Language Models With Activation Engineering2023cited by 99 here
  3. Analyzing the Generalization and Reliability of Steering Vectors2024cited by 33 here
  4. Extracting Latent Steering Vectors from Pretrained Language Models2022cited by 32 here
  5. Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization2024cited by 29 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Extracting Latent Steering Vectors from Pretrained Language Modelscited by 32 here
  2. 2023Steering Language Models With Activation Engineeringcited by 99 here
  3. 2023Steering Llama 2 via Contrastive Activation Additioncited by 159 here
  4. 2024Extending Activation Steering to Broad Skills and Multiple Behaviourscited by 15 here
  5. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 33 here
  6. 2024Programming Refusal with Conditional Activation Steeringcited by 41 here
  7. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 12 here
  8. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 33 here
16 more it builds on, not in this chain
  1. 2022Toy Models of Superpositioncited by 31 here
  2. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 21 here
  3. 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 14 here
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 45 here
  5. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 103 here
  6. 2023Improving Activation Steering in Language Models with Mean-Centringcited by 12 here
  7. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 13 here
  8. 2024Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categoriescited by 18 here
  9. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 29 here
  10. 2024Scaling and evaluating sparse autoencoderscited by 14 here
  11. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 21 here
  12. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 28 here
  13. 2024Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectorscited by 21 here
  14. 2025A Unified Understanding and Evaluation of Steering Methodscited by 19 here
  15. 2025Steering Large Language Model Activations in Sparse Spacescited by 14 here
  16. 2025Understanding (Un)Reliability of Steering Vectors in Language Modelscited by 17 here

The 24 most cited of 235.

Browse

30 of 239, ranked 6–35

  1. 6Improving Instruction-Following in Language Models through Activation Steering202428
  2. 7Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors202421
  3. 8A Unified Understanding and Evaluation of Steering Methods202519
  4. 9Understanding (Un)Reliability of Steering Vectors in Language Models202517
  5. 10Extending Activation Steering to Broad Skills and Multiple Behaviours202415
  6. 11Steering Large Language Model Activations in Sparse Spaces202514
  7. 12Improving Activation Steering in Language Models with Mean-Centring202312
  8. 13Multi-property Steering of Large Language Models with Dynamic Activation Composition202411
  9. 14Towards Reliable Evaluation of Behavior Steering Interventions in LLMs202411
  10. 15Interpretable Steering of Large Language Models with Feature Guided Activation Additions20259
All 30
  1. 16LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss20258
  2. 17HyperSteer: Activation Steering at Scale with Hypernetworks20258
  3. 18Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineering20248
  4. 19Improved Representation Steering for Language Models20257
  5. 20Householder Pseudo-Rotation: A Novel Approach to Activation Editing in LLMs with Direction-Magnitude Perspective20247
  6. 21ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment20266
  7. 22Personalized Text Generation with Contrastive Activation Steering20256
  8. 23Adaptive Multi-Subspace Representation Steering for Attribute Alignment in Large Language Models20256
  9. 24Mechanistic Interpretability of Emotion Inference in Large Language Models20255
  10. 25Who's asking? User personas and the mechanics of latent misalignment20245
  11. 26EmoSteer-TTS: Fine-Grained and Training-Free Emotion-Controllable Text-to-Speech via Activation Steering20254
  12. 27Learning to Steer: Input-dependent Steering for Multimodal LLMs20254
  13. 28DISCO: Disentangled Communication Steering for Large Language Models20254
  14. 29Controlling Large Language Model Agents with Entropic Activation Steering20244
  15. 30In-Distribution Steering: Balancing Control and Coherence in Language Model Generation20253
  16. 31Steering Evaluation-Aware Language Models to Act Like They Are Deployed20253
  17. 32Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering20253
  18. 33AtP*: An efficient and scalable method for localizing LLM behaviour to components20243
  19. 34Do Personality Traits Interfere? Geometric Limitations of Steering in Large Language Models20262
  20. 35The Effectiveness of Style Vectors for Steering Large Language Models: A Human Evaluation20262