Research map / Architectures and efficiency / Pruning and efficient tuning
Zeroth-Order Fine-Tuning of LLMs in Random Subspaces
Ziming Yu, Pan Zhou, Sike Wang, Jia Li, Mi Tian, Hua Huang
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 column-orthogonal projection matrices (ablation) | SubZero · RTE (OPT-13B, FT) | accuracy: 67.5 -> 74.0 | Improved |
| add column-orthogonal projection matrices (ablation) | SubZero · WSC (OPT-13B, FT) | accuracy: 59.6 -> 65.1 | Improved |
| add reshaping of non-square matrices (ablation) | SubZero · SST-2 prompt tuning (OPT-1.3B) | accuracy: 74.2 -> 89.1 | Improved |
| add reshaping of non-square matrices (ablation) | SubZero · SST-2 prefix tuning (OPT-1.3B) | accuracy: 89.4 -> 92.2 | Improved |
| add reshaping of non-square matrices (ablation) | SubZero · SST-2 LoRA (OPT-1.3B) | accuracy: 92.1 -> 92.9 | Improved |
| use random projection matrix instead of activation matrix (ablation) | SubZero · RTE (OPT-13B, FT) | accuracy: 53.1 -> 74.0 | Improved |
| use random projection matrix instead of weight matrix (ablation) | SubZero · RTE (OPT-13B, FT) | accuracy: 65.3 -> 74.0 | Improved |
| update subspace every 1000 steps instead of 20000 (rank 128) (ablation) | SubZero · RTE (OPT-13B, FT) | accuracy: 68.6 -> 74.0 | Improved |
| vary subspace rank r among 32, 64, 128 (ablation) | SubZero · RTE (OPT-13B, FT) | accuracy: 72.6 / 70.0 / 72.2 | No clear effect |
| replace MeZO with SubZero in prompt tuning | MeZO · CB prompt tuning (LLaMA2-7B) | accuracy: 60.7 -> 66.1 | Improved |
| replace MeZO with SubZero in full-parameter tuning | MeZO · CB FT (LLaMA2-7B) | accuracy: 64.3 -> 71.4 | Improved |
| replace S-RGF subspace method with SubZero for memory | S-RGF · SST-2 FT memory (RoBERTa-large) | memory (GB): 23.845 -> 2.690 | Improved |
Closest papers by meaning
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · similarity 0.94)
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 2024 · similarity 0.92)
- 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 0.91)
- Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- A Memory Efficient Randomized Subspace Optimization Method for Training Large Language Models (ICML 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.
