Research map / Large language models / Alignment and preferences
SDA: Steering-Driven Distribution Alignment for Open LLMs Without Fine-Tuning
Wei Xia, Zhi-Hong Deng
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 steering-based logit realignment to the base model (ablation) | Llama2-7B-Chat · Helpfulness (5 datasets) | win rate (%): 57.9% | Improved |
| add steering-based logit realignment to the base model (ablation) | Llama2-7B-Chat · Honesty (TruthfulQA) | win rate (%): 25.9% | Improved |
| add steering-based logit realignment to the base model (ablation) | Llama2-7B-Chat · Harmlessness | win rate (%): 5.3% | Improved |
| add divergence-aware temperature scaling on top of steering (ablation) | SDA (Llama2-7B-Chat) · Helpfulness | win rate (%): 17.9% | Improved |
| add divergence-aware temperature scaling on top of steering (ablation) | SDA (Llama2-7B-Chat) · Honesty | win rate (%): 5.7% | Improved |
| add divergence-aware temperature scaling on top of steering (ablation) | SDA (Llama2-7B-Chat) · Harmlessness | win rate (%): -2.1% | Hurt |
| apply SDA instead of the trained Aligner-7B | Eight open LLMs · 3H, helpfulness | win rate vs Aligner (%): 71.7% | Improved |
| apply SDA instead of the trained Aligner-7B | Eight open LLMs · Honesty (TruthfulQA) | win rate vs Aligner (%): 30.9% | Improved |
| apply SDA instead of the trained Aligner-7B | Open LLMs · Harmlessness | win rate vs Aligner (%): lags behind Aligner-7B in harmlessness by only 1% | No clear effect |
| apply SDA to the base model | Vicuna-13B-V1.5 · E-Dialogue (Empathy) | win rate vs Base (%): 64.5% -> 91.6% (full vs refined) | Improved |
| apply SDA to the base model | DeepSeek-R1-Distill-Qwen-14B · HarmfulQA (Harmless) | win rate vs Base (%): 21.1% -> -0.7% | Mixed |
| apply SDA to the base model | Eight open LLMs · 3H average | win rate vs Base (%): 64.4% helpfulness, 30% honesty, 11.5% harmlessness | Improved |
Closest papers by meaning
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.93)
- Leveraging robust optimization for llm alignment under distribution shifts (NeurIPS 2025 · similarity 0.91)
- Align-Pro: A Principled Approach to Prompt Optimization for LLM Alignment (AAAI 2025 · similarity 0.91)
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.91)
- The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning (ICLR 2024 · similarity 0.91)
- Aligning to Thousands of Preferences via System Message Generalization (NeurIPS 2024 · similarity 0.91)
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.
