Research map / Large language models / Alignment and preferences
Leveraging Machine Unlearning for Cost-Efficient Preference Alignment
XiaoHua Feng, Yuyuan Li, HuWei Ji, Li Zhang, Jiaming Zhang, Tianyu Du, Chaochao Chen
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 U2A sample selection and weighting to gradient ascent unlearning | GA · SafeRLHF, LLaMA2 | Reward-V: -19.103 -> -15.993 | Improved |
| add U2A to GA | GA · SafeRLHF, LLaMA3 | Reward-V: -18.011 -> -7.609 | Improved |
| add U2A to NPO | NPO · SafeRLHF, LLaMA2 | Reward-V: -20.041 -> -19.639 | No clear effect |
| add U2A to GA | GA · SafeRLHF, LLaMA2 | PPL: 7.003 -> 11.021 | Hurt |
| use U2A-weighted unlearning instead of retraining on negatives | GA & U2A · SafeRLHF, LLaMA2 | Reward-V: -18.243 -> -15.993 | Improved |
| use U2A with negatives only vs DPO with positive samples | GA & U2A · UltraFeedback, LLaMA3-8B | Reward-V: -0.277 (DPO) -> 1.105 | Improved |
| use U2A with negatives only vs DPO | GA & U2A · UltraFeedback, LLaMA2-7B | Reward-V: -6.728 (DPO) -> -7.154 | Hurt |
| select Top-K negatives by marginal gain instead of random selection (ablation) | U2A · UltraFeedback, LLaMA3 | average reward | Improved |
| increase number of initially selected samples M beyond a threshold (ablation) | U2A · SafeRLHF/UltraFeedback, LLaMA2/LLaMA3 | average reward | Hurt |
| add positive sample guidance to U2A training (ablation) | U2A · preference alignment | PA performance | Improved |
Closest papers by meaning
- Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning (NeurIPS 2025 · similarity 0.94)
- Large Language Model Unlearning (NeurIPS 2024 · similarity 0.92)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.92)
- WAGLE: Strategic Weight Attribution for Effective and Modular Unlearning in Large Language Models (NeurIPS 2024 · similarity 0.91)
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.91)
- Clean First, Align Later: Benchmarking Preference Data Cleaning for Reliable LLM Alignment (NeurIPS 2025 · 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.
