Research map / Large language models / Alignment and preferences
Sharpness-Aware Minimization in Logit Space Efficiently Enhances Direct Preference Optimization
Haocheng Luo, Zehang Deng, Thanh-Toan Do, Mehrtash Harandi, Dinh Phung, Trung Le
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 logits-SAM (perturb only output layer) | DPO · HH vs chosen | win rate (%): 56.35 -> 60.51 | Improved |
| add logits-SAM | DPO · TL;DR vs SFT | win rate (%): 84.21 -> 89.58 | Improved |
| add logits-SAM | SLiC-HF · HH vs SFT | win rate (%): 65.27 -> 71.87 | Improved |
| add logits-SAM | CPO · TL;DR vs chosen | win rate (%): 39.38 -> 45.41 | Improved |
| add logits-SAM | DPO (Mistral-7B) · Arena-Hard v0.1 | win rate (%): 19.0 -> 23.1 | Improved |
| add logits-SAM | CPO (Mistral-7B) · AlpacaEval 2 LC | LC win rate (%): 8.97 -> 13.32 | Improved |
| add logits-SAM | DPO (Mistral-7B) · MT-Bench | score: 5.49 -> 5.79 | Improved |
| add logits-SAM | DPO (Gemma-2B-IT) · SorryBench test refusal | refusal rate: 0.7077 -> 0.7538 | Improved |
| add logits-SAM | CHES (Gemma-2B-IT) · SorryBench test refusal | refusal rate: 0.7846 -> 0.8769 | Improved |
| increase rho from 1e-4 to 1e-2 (ablation) | logits-SAM DPO · HH vs SFT | win rate (%): 72.28 -> 65.49 | Hurt |
| use rho = 1e-5 instead of 0 (AdamW) (ablation) | logits-SAM DPO · HH vs SFT | win rate (%): 70.52 -> 69.47 (vs chosen 56.35 -> 58.27) | Mixed |
| use logits-SAM instead of vanilla SAM (ablation) | DPO · Pythia-2.8B training cost | wall-clock time: adds ~2-3% time; vanilla SAM nearly doubles step time and OOMs | Improved |
Closest papers by meaning
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · similarity 0.89)
- Dynamics-Aware Preference Optimization for Vision-Language Models (CVPR 2026 · similarity 0.89)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.89)
- Proximalized Preference Optimization for Diverse Feedback Types: A Decomposed Perspective on DPO (NeurIPS 2025 · similarity 0.89)
- Adaptive Batch-Wise Sample Scheduling for Direct Preference Optimization (NeurIPS 2025 · similarity 0.88)
- Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective (ICLR 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.
