Research map / Large language models / Language, safety and interpretability
Steering at the Source: Style Modulation Heads for Robust Persona Control
Yoshihiro Izawa, Gouki Minegishi, Koshi Eguchi, Sosuke Hosokawa, Kenjiro Taura
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 |
|---|---|---|---|
| steer only Style Modulation Heads instead of MLP residual stream | Qwen2.5-7B humorous, Neutral+alpha · Pareto score tau=80 | trait score at coherency>=80: 17.8 -> 60.7 | Improved |
| steer only Style Modulation Heads instead of Attn Residual | Qwen2.5-7B humorous, Neutral+alpha · Pareto score tau=80 | trait score at coherency>=80: 29.6 -> 60.7 | Improved |
| steer only Style Modulation Heads instead of Attn Residual | Llama-3.1-8B evil, Neutral+alpha · Pareto score tau=80 | trait score at coherency>=80: 41.4 -> 29.4 | Hurt |
| steer only Style Modulation Heads instead of Attn Residual | Llama-3.1-8B hallucination, Target-alpha · Pareto score tau=80 | trait score at coherency>=80: 50.5 -> 91.2 | Improved |
| steer only Style Modulation Heads instead of Attn Residual | Qwen2.5-7B hallucination, Target-alpha · Pareto score tau=80 | trait score at coherency>=80: 42.8 -> 73.1 | Improved |
| also steer anti-correlated heads (ablation) | Style Modulation Head steering, Qwen2.5-7B hallucination Neutral+alpha · Pareto score tau=80 | trait score at coherency>=80: 69.0 -> 36.5 | Hurt |
| also steer anti-correlated heads (ablation) | Style Modulation Head steering, Llama-3.1-8B evil Target-alpha · Pareto score tau=80 | trait score at coherency>=80: 99.9 -> 13.2 | Hurt |
| also steer anti-correlated heads (ablation) | Style Modulation Head steering, Llama-3.1-8B passionate Neutral+alpha · Pareto score tau=80 | trait score at coherency>=80: 94.5 -> 55.0 | Hurt |
| zero-ablate high-contribution heads in layers 20, 15, 19 (ablation) | Qwen2.5-7B persona expression · trait score | trait score | Improved |
| zero-ablate high-contribution heads (ablation) | Qwen2.5-7B · coherency and MMLU | coherency, MMLU | No clear effect |
| steer the Style Modulation Heads (Head Cor) vs other positions across personas | Qwen2.5-7B and Llama-3.1-8B · 12 conditions per model | highest Pareto score count: 11 cases for Qwen2.5-7B and 9 cases for Llama-3.1-8B | Improved |
| increase steering magnitude in the positive direction | residual stream steering · coherency | coherency score | Hurt |
Closest papers by meaning
- Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.86)
- Head Pursuit: Probing Attention Specialization in Multimodal Transformers (NeurIPS 2025 · similarity 0.86)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.86)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.86)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.85)
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.85)
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.
