Research map / Large language models / Language, safety and interpretability
Towards Understanding Steering Strength
Magamed Taimeskhanov, Samuel Vaiter, Damien Garreau
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 |
|---|---|---|---|
| increase steering strength alpha along a difference-of-means direction (ablation) | next-token probability of concept tokens · GPT-2, Gemma 3, Qwen 3 and others, concepts evil/joy/depression | next-token probability increase | Mixed |
| increase steering strength alpha (ablation) | concept presence in generations · 11 LLMs, judge Gemma 3 12B | concept probability | Improved |
| increase |alpha| away from 0 (ablation) | cross-entropy of steered model · fineweb, 10^6 tokens, all models | cross-entropy increase | Hurt |
| very large alpha (ablation) | cross-entropy of steered model · fineweb | cross-entropy | No clear effect |
| steer early layers instead of mid to late layers (ablation) | activation steering · multiple models and concepts | concept relatedness of top tokens and theory fit | Hurt |
| use only 10 prompt pairs instead of 500 to build the steering vector (ablation) | difference-of-means steering · low-data setting | steering effectiveness | No clear effect |
| use a base model (Qwen3-0.6B-Base) as concept judge instead of instruction-tuned judges (ablation) | concept-presence estimation · judge robustness | concept probability curve | Hurt |
| use a more precise judge prompt with negative examples (ablation) | concept-presence judge for evil on Gemma 3 1B · concept probability | spurious early spike | Improved |
Closest papers by meaning
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.90)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.90)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.90)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.90)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.89)
- Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024 · 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.
