Research map / Large language models / Language, safety and interpretability
Endogenous Resistance to Activation Steering in Language Models
Alex McKenzie, Keenan Pepper, Stijn Servaes, Martin Leitgab, Murat Cubuktepe, Michael Vaiana, Diogo S de Lucena, Judd Rosenblatt, Michael S. A. Graziano
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| append self-monitoring meta-prompt to the prompt (ablation) | Llama-3.3-70B · Steered ESR | multi-attempt rate: 7.4% to 31.7% | Improved |
| append self-monitoring meta-prompt to the prompt (ablation) | Llama-3.3-70B · Steered ESR | ESR rate: 3.8% to 14.8% | Improved |
| keep the self-correction-associated latents active (versus zero-ablating them) (ablation) | Llama-3.3-70B · Steered ESR | ESR rate: 3.8% to 2.8% | Improved |
| keep the self-correction-associated latents active (versus zero-ablating them) (ablation) | Llama-3.3-70B · Steered ESR | multi-attempt rate: 7.4% to 5.5% | Improved |
| zero-ablate the latents (effect on first-attempt quality) (ablation) | Llama-3.3-70B · Steered ESR | mean first-attempt score: 26.3 -> 27.4 | No clear effect |
| ablate random matched latents instead of the targeted ones (ablation) | Llama-3.3-70B · Steered ESR | ESR rate: 3.8% to 4.2% | No clear effect |
| zero-ablate the latents on held-out prompts (ablation) | Llama-3.3-70B · Held-out prompts (20 prompts) | ESR rate: reduces it by 30-45% | Improved |
| add a length-matched on-topic prefix before steered continuation (ablation) | Llama-3.3-70B · Matched-prefix (447 episodes) | continuation relevance score: 8.7 -> 30.4 (+21.8) | Improved |
| use natural post-correction text rather than a generic on-topic prefix (ablation) | Llama-3.3-70B · Matched-prefix (447 episodes) | continuation relevance score: 30.4 -> 54.5; +45.8 vs +21.8 | Improved |
| prefill off-topic text into an unsteered model rather than steering (ablation) | Llama-3.3-70B · Prefilling control | self-correction success: ~50% -> ~95% | Improved |
| fine-tune on more masked self-correction data (ablation) | Llama-3.1-8B · Steered ESR | correction improvement rate | No clear effect |
| fine-tune on higher ratio of self-correction data (10% to 90%) (ablation) | Llama-3.1-8B · Steered ESR | multi-attempt rate | Improved |
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.90)
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.89)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.88)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.88)
- Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors (ICLR 2025 · similarity 0.87)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
