Research map / Large language models / Instruction and fine-tuning
Faithful Bi-Directional Model Steering via Distribution Matching and Distributed Interchange Interventions
Yuntai Bao, Xuhong Zhang, Jintao Chen, Ge Su, Yuxiang Cai, Hao Peng, SUN Bing, Haiqin Weng, Liu Yan, Jianwei Yin
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 |
|---|---|---|---|
| use JSD distribution matching instead of reverse KL (ablation) | CDAS · AxBench Concept500 (Gemma-2-2B L20, 2 concepts) | steering score: 0.840 -> 0.860 | Improved |
| use JSD distribution matching instead of forward KL (ablation) | CDAS · AxBench Concept500 (Gemma-2-2B L20, 2 concepts) | steering score: 0.560 -> 0.860 | Improved |
| tune the steering factor at inference instead of unit factor (ablation) | CDAS · AxBench Gemma-2-2B L10 | steering score: 0.121 -> 0.631 | Improved |
| tune the steering factor at inference instead of unit factor (ablation) | CDAS · AxBench Gemma-2-9B L20 | steering score: 0.238 -> 0.992 | Improved |
| scale model from Gemma-2-2B to Gemma-2-9B (layer 20) | CDAS · AxBench Concept500 | steering score: 0.608 -> 0.992 | Improved |
| use CDAS instead of LoReFT | Gemma-2-9B L20 · AxBench Concept500 | steering score: 0.777 -> 0.992 | Improved |
| use CDAS instead of RePS | Gemma-2-9B L31 · AxBench Concept500 | steering score: 0.624 -> 0.518 | Hurt |
| move intervention from layer 10 to layer 20 | CDAS · AxBench Gemma-2-2B | steering score: 0.631 -> 0.608 | No clear effect |
| move intervention from layer 10 to layer 20 | RePS · AxBench Gemma-2-2B | steering score: 0.756 -> 0.606 | Hurt |
| use bi-directional CDAS instead of DAS objective | Llama-3.1-8B · Refusal override suppression harmonic mean | harmonic mean (%): 2 -> 95 | Improved |
| use CDAS instead of RePS | Phi-3.5-mini · Refusal override suppression harmonic mean | harmonic mean (%): 91 -> 46 | Hurt |
| use CDAS instead of RePS | Llama-3.1-8B · MMLU change when overriding refusal | accuracy change (%): -35.57 -> 0.20 | Improved |
Closest papers by meaning
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.87)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.87)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.87)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.86)
- Angular Steering: Behavior Control via Rotation in Activation Space (NeurIPS 2025 · similarity 0.86)
- Model Steering: Learning with a Reference Model Improves Generalization Bounds and Scaling Laws (ICML 2025 · 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.
