AI Paper Atlas

Research map / Optimization / Optimizers

LOZO+: Provably Efficient Zeroth-Order Fine-Tuning via Greedy Low-Rank Subspace Selection

Jinjie Fang, Chengxun Jin, Tianxing Man, Yi Chang, Bin Gu

ICML 2026 · Optimization › Optimizers · approach: Zeroth-order training methods

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
keep selected directions with greedy refresh (vs no subspace refresh) (ablation)LOZO+ · Average of CB, SQuAD, SST2, WICaccuracy: 81.7 -> 85.7Improved
retain previously selected directions (vs full reinitialization at fixed intervals) (ablation)LOZO+ · Average of CB, SQuAD, SST2, WICaccuracy: 83.0 -> 85.7Improved
add greedy selection vs no refresh on WIC (OPT-13B) (ablation)LOZO+ · WIC (OPT-13B)accuracy: 58.5 -> 64.9Improved
add greedy selection vs full reinit on WIC (OPT-13B) (ablation)LOZO+ · WIC (OPT-13B)accuracy: 59.1 -> 64.9Improved
add greedy selection vs no refresh on CB (Qwen2.5-7B) (ablation)LOZO+ · CB (Qwen2.5-7B)accuracy: 89.3 -> 96.4Improved
replace random LOZO subspace with greedy subspace selection (LOZO+)LOZO · Qwen2.5-7B, 11 tasks averageaccuracy: 82.2 -> 83.3Improved
replace random LOZO subspace with greedy subspace selection (LOZO+)LOZO · OPT-13B, 11 tasks averageaccuracy: 70.5 -> 72.3Improved
replace random LOZO subspace with greedy subspace selection (LOZO+)LOZO · Qwen2.5-32B, 6 tasks averageaccuracy: 85.3 -> 86.8Improved
replace random LOZO subspace with greedy subspace selection (LOZO+)LOZO · OPT-66B, 6 tasks averageaccuracy: 74.7 -> 76.0Improved
LOZO+ vs MeZOMeZO · Qwen2.5-7B, 11 tasks averageaccuracy: 80.9 -> 83.3Improved
LOZO+ vs first-order FT (Adam)FT · Qwen2.5-7B, 11 tasks averageaccuracy: 85.3 -> 83.3Hurt
LOZO+ vs LOZO on ReCoRD-style generation DROP (Qwen2.5-7B)LOZO · DROP (Qwen2.5-7B)accuracy: 60.7 -> 60.8No clear effect

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