Research map / Architectures and efficiency / Pruning and efficient tuning
Low-Rank Curvature for Zeroth-Order Optimization in LLM Fine-tuning
Hyunseok Seung, Jaewoo Lee, Hyunsuk Ko
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 RLOO leave-one-out baseline to ZO-SGD (ablation) | ZO-SGD · Sphere, Rastrigin, Rosenbrock (1,000-dim) | gradient estimate MSE | Improved |
| replace MeZO with LOREN | MeZO · DistilBERT GLUE avg | accuracy (%): 54.4 -> 65.4 | Improved |
| replace MeZO-Adam with LOREN | MeZO-Adam · RoBERTa-large GLUE avg | accuracy (%): 73.6 -> 70.1 | Hurt |
| replace MeZO-Adam with LOREN | MeZO-Adam · GPT-2-XL GLUE avg | accuracy (%): 71.5 -> 71.9 | Improved |
| replace MeZO with LOREN | MeZO · OPT-2.7B GLUE avg | accuracy (%): 70.7 -> 71.6 | Improved |
| replace MeZO with LOREN | MeZO · LLaMA-3-8B SuperGLUE avg | accuracy (%): 63.6 -> 63.7 | No clear effect |
| replace MeZO with LOREN | MeZO · OPT-13B SuperGLUE avg | accuracy (%): 63.4 -> 63.5 | No clear effect |
| use LOREN instead of HiZOO | HiZOO · OPT-2.7B GLUE avg | accuracy (%): 70.0 -> 71.6 | Improved |
| use six forward passes per step instead of two or three | ZO optimizers · GLUE and SuperGLUE | accuracy | Improved |
| use LOREN instead of MeZO-Adam for memory | MeZO-Adam · OPT-13B peak memory (GB) | memory: 76.0 -> 57.5 | Improved |
| use LOREN instead of MeZO | MeZO · OPT-13B peak memory (GB) | memory: 32.9 -> 57.5 | Hurt |
| use LOREN instead of MeZO for query efficiency | MeZO · GPT-2-XL SST-2 forward passes to 70% | # queries: 16,752 -> 1,320 | Improved |
Closest papers by meaning
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · 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)
- Second-Order Fine-Tuning without Pain for LLMs: A Hessian Informed Zeroth-Order Optimizer (ICLR 2025 · similarity 0.92)
- Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning (NeurIPS 2025 · similarity 0.90)
- Bilevel ZOFO: Efficient LLM Fine-Tuning and Meta-Training (NeurIPS 2025 · similarity 0.90)
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.
