mlsanityData to 14 Oct 2026

Steering and probing for LLM safety

119 papers0 this weekMedium confidence in this grouping: mostly one topic, with blurry edgesMedium confidence in this grouping: mostly one topic, with blurry edges · About

Keywords: safety, refusal, steering, harmful, llms

Key facts. Steering and probing for LLM safety: 119 papers on arXiv since 2023 (data through 14 Oct 2026). 17 of its 119 mapped papers (14%) cite “Programming Refusal with Conditional Activation Steering” (2024, arXiv 2409.05907). 10 new papers 24 Sep–14 Oct.

What’s new

24 Sep – 14 Oct 2026

10 new papers, 24 Sep – 14 Oct; 0 this week.

  1. Earlier · 10
  2. Nullify: Null-Space Activation Steering for Training-Free LLM Unlearning also fits: Activation steering methods
  3. SafeEvo: Deciphering the Safety Alignment Mechanism and Evolution in Language Models
  4. Backdooring Sparse Autoencoders also fits: SAE features and their applications
  5. Don't Judge an LLM Only by Its Activations: Discovering Suppressed Safety Features via Counterfactual Activation Potential
  6. EmoRSS: Mitigating Emotion-Induced Over-Refusal in Large Language Models
  7. Detect and Suppress: A Mechanistic Defense against Adversarial Patches in VLA Models
  8. CounterSteer: Suppressing Indirect Prompt Injection with Activation Steering
  9. Less Sycophancy, Stronger Refusal? Lessons for AI Safety from Mechanistic Interpretability
  10. When Do Model Internals Help? Exploring the Role of Representation Engineering in LLM Safety
  11. STR: Supervised Transcoder Replacement for Reducing Steering Side Effects

Foundations

Most cited by papers here

  1. Programming Refusal with Conditional Activation Steering2024cited by 17 here
  2. Steering Language Model Refusal with Sparse Autoencoders2024cited by 14 here
  3. Steering Without Side Effects: Improving Post-Deployment Control of Language Models2024cited by 8 here
  4. The Rogue Scalpel: Activation Steering Compromises LLM Safety2025cited by 6 here
  5. Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks2024cited by 5 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 14 here
  2. 2023Progress measures for grokking via mechanistic interpretabilitycited by 5 here
  3. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 8 here
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 39 here
  5. 2024Scaling and evaluating sparse autoencoderscited by 12 here
  6. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 11 here
  7. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 17 here
  8. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 9 here
  9. 2025The Rogue Scalpel: Activation Steering Compromises LLM Safetycited by 6 here
15 more it builds on, not in this chain
  1. 2022Extracting Latent Steering Vectors from Pretrained Language Modelscited by 10 here
  2. 2022Toy Models of Superpositioncited by 8 here
  3. 2023Steering Language Models With Activation Engineeringcited by 30 here
  4. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 44 here
  5. 2023Steering Llama 2 via Contrastive Activation Additioncited by 49 here
  6. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 13 here
  7. 2024Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categoriescited by 6 here
  8. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 6 here
  9. 2024Steering Without Side Effects: Improving Post-Deployment Control of Language Modelscited by 8 here
  10. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 8 here
  11. 2024Programming Refusal with Conditional Activation Steeringcited by 17 here
  12. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 10 here
  13. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 5 here
  14. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 14 here
  15. 2025Open Problems in Mechanistic Interpretabilitycited by 6 here

The 24 most cited of 193.

Browse

30 of 119, ranked 6–35

  1. 6Understanding Refusal in Language Models with Sparse Autoencoders20255
  2. 7AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint20254
  3. 8Mechanistic interpretability for steering vision-language-action models20254
  4. 9Revisiting Jailbreaking for Large Language Models: A Representation Engineering Perspective20244
  5. 10Trojan Activation Attack: Red-Teaming Large Language Models using Activation Steering for Safety-Alignment20234
  6. 11Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models20263
  7. 12Sparse Autoencoders are Capable LLM Jailbreak Mitigators20263
  8. 13AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender20253
  9. 14Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoders20253
  10. 15Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusal20253
All 30
  1. 16Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons20243
  2. 17There Is More to Refusal in Large Language Models than a Single Direction20262
  3. 18One-shot Optimized Steering Vectors Mediate Safety-relevant Behaviors in LLMs20252
  4. 19Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework20252
  5. 20Backdoor Attribution: Elucidating and Controlling Backdoor in Language Models20252
  6. 21SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering20242
  7. 22Towards Inference-time Category-wise Safety Steering for Large Language Models20242
  8. 23Analysing the Safety Pitfalls of Steering Vectors20261
  9. 24Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control20261
  10. 25Contrastive Conceptor Activation Steering (COAST): Unlocking Vision-Language-Action Models through Hidden States20261
  11. 26Using Mechanistic Interpretability to Craft Adversarial Attacks against Large Language Models20251
  12. 27Guiding Giants: Lightweight Controllers for Weighted Activation Steering in LLMs20251
  13. 28SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMs20251
  14. 29Beyond Linear Probes: Dynamic Safety Monitoring for Language Models20251
  15. 30SteerVLM: Robust Model Control through Lightweight Activation Steering for Vision Language Models20251
  16. 31Graph-Regularized Sparse Autoencoders for LLM Safety Steering20251
  17. 32Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics20251
  18. 33No Two Devils Alike: Unveiling Distinct Mechanisms of Fine-tuning Attacks20241
  19. 34What Features in Prompts Jailbreak LLMs? Investigating the Mechanisms Behind Attacks20241
  20. 35Linear Probe Penalties Reduce LLM Sycophancy20241