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
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 |
|---|---|---|---|
| keep selected directions with greedy refresh (vs no subspace refresh) (ablation) | LOZO+ · Average of CB, SQuAD, SST2, WIC | accuracy: 81.7 -> 85.7 | Improved |
| retain previously selected directions (vs full reinitialization at fixed intervals) (ablation) | LOZO+ · Average of CB, SQuAD, SST2, WIC | accuracy: 83.0 -> 85.7 | Improved |
| add greedy selection vs no refresh on WIC (OPT-13B) (ablation) | LOZO+ · WIC (OPT-13B) | accuracy: 58.5 -> 64.9 | Improved |
| add greedy selection vs full reinit on WIC (OPT-13B) (ablation) | LOZO+ · WIC (OPT-13B) | accuracy: 59.1 -> 64.9 | Improved |
| add greedy selection vs no refresh on CB (Qwen2.5-7B) (ablation) | LOZO+ · CB (Qwen2.5-7B) | accuracy: 89.3 -> 96.4 | Improved |
| replace random LOZO subspace with greedy subspace selection (LOZO+) | LOZO · Qwen2.5-7B, 11 tasks average | accuracy: 82.2 -> 83.3 | Improved |
| replace random LOZO subspace with greedy subspace selection (LOZO+) | LOZO · OPT-13B, 11 tasks average | accuracy: 70.5 -> 72.3 | Improved |
| replace random LOZO subspace with greedy subspace selection (LOZO+) | LOZO · Qwen2.5-32B, 6 tasks average | accuracy: 85.3 -> 86.8 | Improved |
| replace random LOZO subspace with greedy subspace selection (LOZO+) | LOZO · OPT-66B, 6 tasks average | accuracy: 74.7 -> 76.0 | Improved |
| LOZO+ vs MeZO | MeZO · Qwen2.5-7B, 11 tasks average | accuracy: 80.9 -> 83.3 | Improved |
| LOZO+ vs first-order FT (Adam) | FT · Qwen2.5-7B, 11 tasks average | accuracy: 85.3 -> 83.3 | Hurt |
| LOZO+ vs LOZO on ReCoRD-style generation DROP (Qwen2.5-7B) | LOZO · DROP (Qwen2.5-7B) | accuracy: 60.7 -> 60.8 | No clear effect |
Closest papers by meaning
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · similarity 0.97)
- Variance-reduced Zeroth-Order Methods for Fine-Tuning Language Models (ICML 2024 · similarity 0.92)
- Harmony in Divergence: Towards Fast, Accurate, and Memory-efficient Zeroth-order LLM Fine-tuning (NeurIPS 2025 · similarity 0.92)
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 2024 · similarity 0.92)
- Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning (NeurIPS 2025 · similarity 0.92)
- Dominant-Layer ZO: A Single Layer Dominates Zeroth-Order Fine-Tuning of LLMs (NeurIPS 2026 · 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.
