AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning

Yicheng Lang, Yihua Zhang, Chongyu Fan, Changsheng Wang, Jinghan Jia, Sijia Liu

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

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
replace Adam with zeroth-order AdaZO optimizer (ablation)NPO unlearning · MUSE-Books, 4-bit quantizationVerbMem on forget set after quantization: 48.5 -> 16.2Improved
replace Adam with zeroth-order AdaZO optimizer (ablation)NPO unlearning · MUSE-Books, no quantization or attackKnowMem on retain set: 53.65 -> 36.6Hurt
replace Adam with zeroth-order AdaZO optimizer (ablation)NPO unlearning · MUSE-Books, before quantizationVerbMem on forget set: 0 -> 20.7Hurt
replace Adam with signSGD or signAdam gradient compression (ablation)NPO unlearning · MUSE-Books, 4-bit quantizationVerbMem on forget set after quantization: 48.5 -> 15.6 (signSGD), 30.7 (signAdam)Improved
replace Adam with signSGD or signAdam (ablation)unlearning · MUSE-Books, relearningrobustness to relearningNo clear effect
replace Adam with first-order randomized smoothing (RS) objective (ablation)NPO unlearning · MUSE-Books, 4-bit quantizationVerbMem on forget set after quantization: 48.5 -> 34.5Improved
add Hybrid FO-ZO alternation instead of Adam onlyNPO unlearning · MUSE-Books, 4-bit quantizationVerbMem on forget set after quantization: 48.5 -> 0Improved
add Hybrid FO-ZO alternation instead of Adam onlyNPO unlearning · MUSE-Books, relearning attackKnowMem on forget set after relearning: 34.8 -> 0Improved
replace SAM with Hybrid FO-ZONPO unlearning · MUSE-Books, 4-bit quantizationVerbMem on forget set after quantization: 30.0 -> 0Improved
vary hybrid switch step N between FO and ZO (5, 20, 100) (ablation)Hybrid optimizer · MUSE-Books, NPOrobustness to quantization and relearningNo clear effect
assign more FO than ZO steps in each hybrid round (FO=20, ZO=10) (ablation)Hybrid optimizer · MUSE-Books, NPOunlearning robustnessHurt
assign more ZO than FO steps in each hybrid round (FO=10, ZO=20) (ablation)Hybrid optimizer · MUSE-Books, NPOunlearning effectivenessHurt

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