Research map / Architectures and efficiency / Pruning and efficient tuning
CurvZO: Adaptive Curvature-Guided Sparse Zeroth-Order Optimization for Efficient LLM Fine-Tuning
Shuo Wang, Ziyu Chen, Ming Tang
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 curvature-guided sparse perturbation instead of dense MeZO | MeZO · OPT-2.7B, SuperGLUE/generation average | accuracy (%): 64.2 -> 66.8 | Improved |
| use CurvZO instead of MeZO on WSC | MeZO · OPT-2.7B, WSC | accuracy (%): 58.6 -> 62.9 | Improved |
| use CurvZO instead of MeZO on SQuAD | MeZO · OPT-2.7B, SQuAD | accuracy (%): 79.0 -> 81.5 | Improved |
| use CurvZO instead of MeZO on SQuAD | MeZO · OPT-6.7B, SQuAD | accuracy (%): 81.6 -> 83.7 | Improved |
| use CurvZO instead of MeZO on WIC | MeZO · OPT-6.7B, WIC | accuracy (%): 62.5 -> 61.7 | Hurt |
| use CurvZO instead of MeZO with LoRA | MeZO + LoRA · OPT-2.7B, average | accuracy (%): 62.1 -> 65.1 | Improved |
| use CurvZO instead of MeZO | MeZO · OPT-2.7B, SST-2 | accuracy (%): 92.1 -> 94.1 | Improved |
| use CurvZO instead of DiZO | DiZO · OPT-2.7B, SST-2 | accuracy (%): 94.4 -> 94.1 | Hurt |
| add LoRA to CurvZO | CurvZO · OPT-2.7B, average | accuracy (%): 66.8 -> 65.1 | Hurt |
| use CurvZO instead of MeZO for convergence speed | MeZO · OPT-2.7B, BoolQ | optimization steps to same accuracy: approximately 2.4× fewer steps | Improved |
| use CurvZO instead of MeZO for GPU hours | MeZO · OPT-2.7B, BoolQ | GPU hours: -59% | Improved |
| store curvature scores | MeZO · OPT-6.7B | memory (GB): 13.94 -> 13.95 | No clear effect |
Closest papers by meaning
- 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 0.91)
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · similarity 0.90)
- Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning (NeurIPS 2026 · similarity 0.89)
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.
