Research map / Large language models / Language, safety and interpretability
Steer Like the LLM: Activation Steering that Mimics Prompting
Geert Heyman, Frederik Vandeputte
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 |
|---|---|---|---|
| add token-specific steering coefficients (PSR) to single-layer MSE steering (ablation) | S-Const MSE|QR · Persona Vectors, Llama-3.2-3B | TA@Cp: 57.4 -> 66.8 | Improved |
| add token-specific steering coefficients (PSR) to single-layer LL steering (ablation) | S-Const LL|QR · Persona Vectors, Llama-3.2-3B | TA@Cp: 42.9 -> 52.6 | Improved |
| add token-specific coefficients (PSR) to all-layer MSE steering (ablation) | A-Const MSE|QR · Persona Vectors, Llama-3.2-3B | TA@Cp: 95.8 -> 92.5 | Hurt |
| add token-specific coefficients (PSR) to all-layer MSE steering (ablation) | A-Const MSE|QR · Persona Vectors, Llama-3.1-8B | TA@Cp: 91.3 -> 96.4 | Improved |
| use MSE instead of loglikelihood objective (ablation) | S-PSR|QR · Persona Vectors, Llama-3.1-8B | TA@Cp: 45.0 -> 74.7 | Improved |
| use loglikelihood instead of MSE objective (ablation) | S-PSR · IFEval format, Gemma-2-9b-it | IF accuracy: 47.5 -> 66.1 | Improved |
| add token-specific coefficients (PSR) to single-layer LL steering (ablation) | S-Const LL · IFEval format, Gemma-2-9b-it | IF accuracy: 13.4 -> 66.1 | Improved |
| steer all layers instead of a single layer (ablation) | PSR LL · IFEval format, Gemma-2-9b-it | IF accuracy: 66.1 -> 71.9 | Mixed |
| use PSR MSE instead of constant MSE on AxBench (ablation) | A-Const MSE · AxBench Gemma-2-2B L20 | steering score: 0.783 -> 0.871 | Improved |
| use PSR MSE instead of constant MSE on AxBench (ablation) | A-Const MSE · AxBench Gemma-2-9B L20 | steering score: 1.053 -> 1.120 | Improved |
| use PSR LL instead of constant LL on AxBench (ablation) | A-Const LL · AxBench Gemma-2-2B L20 | steering score: 0.792 -> 0.690 | Hurt |
| use MSE instead of LL for single-layer PSR on AxBench (ablation) | S-PSR · AxBench Gemma-2-9B L20 | steering score: 0.667 -> 0.900 | Improved |
Closest papers by meaning
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.90)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.89)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.89)
- Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs (NeurIPS 2026 · similarity 0.89)
- Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.88)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.88)
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.
