AI Paper Atlas

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

ICML 2026 · Large language models › Alignment and preferences · approach: Robust and calibrated DPO · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add U2A sample selection and weighting to gradient ascent unlearningGA · SafeRLHF, LLaMA2Reward-V: -19.103 -> -15.993Improved
add U2A to GAGA · SafeRLHF, LLaMA3Reward-V: -18.011 -> -7.609Improved
add U2A to NPONPO · SafeRLHF, LLaMA2Reward-V: -20.041 -> -19.639No clear effect
add U2A to GAGA · SafeRLHF, LLaMA2PPL: 7.003 -> 11.021Hurt
use U2A-weighted unlearning instead of retraining on negativesGA & U2A · SafeRLHF, LLaMA2Reward-V: -18.243 -> -15.993Improved
use U2A with negatives only vs DPO with positive samplesGA & U2A · UltraFeedback, LLaMA3-8BReward-V: -0.277 (DPO) -> 1.105Improved
use U2A with negatives only vs DPOGA & U2A · UltraFeedback, LLaMA2-7BReward-V: -6.728 (DPO) -> -7.154Hurt
select Top-K negatives by marginal gain instead of random selection (ablation)U2A · UltraFeedback, LLaMA3average rewardImproved
increase number of initially selected samples M beyond a threshold (ablation)U2A · SafeRLHF/UltraFeedback, LLaMA2/LLaMA3average rewardHurt
add positive sample guidance to U2A training (ablation)U2A · preference alignmentPA performanceImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy