Research map / Large language models / Instruction and fine-tuning
Towards Steering without Sacrifice: Principled Training of Steering Vectors for Prompt-only Interventions
Yuntai Bao, Qinfeng Li, Xinyan Yu, Ge Su, Wenqi Zhang, Liu Yan, Haiqin Weng, Jianwei Yin, Xuhong Zhang
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 |
|---|---|---|---|
| train steering factor jointly with direction instead of fixed unit factor (ablation) | FSSV (Lang.) · AxBench CONCEPT500 | overall steering score: 0.072 -> 0.736 (Gemma2-2B) | Improved |
| add joint training instead of factor sampling plus selection (ablation) | FSSV (Lang.) · AxBench CONCEPT500 | overall steering score: 0.663 -> 0.736 (Gemma2-2B) | Improved |
| add joint training to SimPO FSSV (RePS) on Gemma2-9B (ablation) | FSSV (SimPO) · AxBench CONCEPT500 (Gemma2-9B) | overall steering score: 0.892 -> 0.886 | No clear effect |
| use prompt-only intervention (PrOSV) instead of full-sequence (ablation) | SV with Lang. objective · AxBench CONCEPT500 (Gemma2-2B) | overall steering score: 0.736 -> 0.758 | Improved |
| use prompt-only intervention (PrOSV) instead of full-sequence with SimPO (ablation) | SV with SimPO objective · AxBench (Qwen2.5-32B) | overall steering score: 0.982 -> 1.102 | Improved |
| use SimPO objective instead of language modeling (ablation) | PrOSV · AxBench (Gemma2-2B) | overall steering score: 0.758 -> 0.803 | Improved |
| use larger factor initialization size and factor learning rate (ablation) | joint SV training · CONCEPT10 | concept score | Improved |
| increase direction initialization size lambda from 1 to 8 (ablation) | FSSV · CONCEPT10 | concept score | No clear effect |
| increase direction initialization size lambda from 1 to 8 (ablation) | PrOSV · CONCEPT10 | concept score | Improved |
| intervene on full prompt instead of few tokens (ablation) | PrOSV · CONCEPT10 (Gemma2-9B) | overall score: 0.92 -> 0.71 (overall); concept 1.14 -> 1.41 | Hurt |
| use prefix-suffix p2+s2 instead of p1+s1 or p4+s4 (ablation) | PrOSV · CONCEPT10 (Qwen2.5-32B) | overall score: 1.10 -> 1.16 (p1+s1 to p2+s2) | Improved |
| use PrOSV instead of FSSV for utility | Gemma2-9B steering · tinyGSM8K | accuracy %: 8.6 -> 68.4 | Improved |
Closest papers by meaning
- Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024 · similarity 0.92)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.89)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.89)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.89)
- Model Whisper: Steering Vectors Unlock Large Language Models’ Potential in Test-Time (AAAI 2026 · similarity 0.88)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · 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.
