Research map / Architectures and efficiency / Pruning and efficient tuning
FZOO: Fast Zeroth-Order Optimizer for Fine‑Tuning Large Language Models towards Adam‑Scale Speed
Sizhe Dang, yangyangGuo, Yanjun Zhao, Xiaodong Zheng, Guang Dai, Ivor Tsang, Haishan Ye
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 |
|---|---|---|---|
| use FZOO instead of MeZO | RoBERTa-large · 6 NLP tasks, k=16 | average accuracy: 67.4 -> 71.4 | Improved |
| use FZOO instead of MeZO | Phi-2, Llama3, OPT-13B · 11 tasks | average accuracy: 70.7 -> 73.0 (Phi-2); 75.4 -> 77.2 (Llama3); 68.8 -> 70.7 (OPT-13B) | Improved |
| use FZOO instead of MeZO | OPT-30B / OPT-66B · SST-2, RTE, WSC, WIC | average accuracy: 69.2 -> 71.1 (30B); 69.8 -> 71.5 (66B) | Improved |
| use FZOO with non-differentiable F1 objective | OPT · SQuAD | F1: 70.5 -> 74.4 | Improved |
| add adaptive step-size and batched one-sided estimates | FZOO · RoBERTa-large loss convergence | forward passes to converge: 18x fewer forward passes than MeZO | Improved |
| use FZOO vs MeZO | FZOO · ReCoRD and DROP | accuracy: DROP 57.1 -> 56.5 (Llama3) | Hurt |
| combine FZOO with prefix-tuning | FZOO(prefix) · RoBERTa-large, k=16 | average accuracy: 69.6 -> 70.3 vs MeZO(prefix) | Improved |
| use FZOO instead of ZO-Adam | FZOO · SST-2/COPA multi-model | average accuracy: 86.1 -> 90.4 | Improved |
| increase perturbations N per step (ablation) | FZOO · SST-2, OPT-125M | accuracy | Mixed |
| use FZOO per-step compute vs MeZO | FZOO · RoBERTa-large | wall-clock per step: 0.4305s -> 0.6052s | Hurt |
Closest papers by meaning
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 2024 · similarity 0.91)
- Addax: Utilizing Zeroth-Order Gradients to Improve Memory Efficiency and Performance of SGD for Fine-Tuning Language Models (ICLR 2025 · similarity 0.91)
- Towards Efficient Low-Order Hybrid Optimizer for Language Model Fine-Tuning (AAAI 2025 · similarity 0.91)
- Second-Order Fine-Tuning without Pain for LLMs: A Hessian Informed Zeroth-Order Optimizer (ICLR 2025 · similarity 0.91)
- Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation is Wasteful (NeurIPS 2025 · similarity 0.90)
- Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training (ICLR 2024 · 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.
