Research map / Architectures and efficiency / Pruning and efficient tuning
TeZO: Empowering the Low-Rankness on the Temporal Dimension in the Zeroth-Order Optimization for Fine-Tuning LLMs
Yan Sun, Qixin Zhang, Liang Ding, Shen Li, Dacheng Tao
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 Adam-style adaptivity with low-rank temporal structure (TeZO-Adam vs TeZO) (ablation) | TeZO · OPT-13B | average gap vs FT: -5.2 -> -2.5 | Improved |
| add momentum (TeZO-m vs TeZO) (ablation) | TeZO · OPT-13B | average gap vs FT: -5.2 -> -4.8 | Improved |
| use TeZO-Adam instead of MeZO-Adam | TeZO-Adam · OPT-13B | average gap vs FT: -3.2 -> -2.5 | Improved |
| use TeZO-Adam instead of MeZO-Adam | TeZO-Adam · LLaMA-7B | average gap vs FT: -7.5 -> -7.0 | Improved |
| use TeZO instead of MeZO | TeZO · LLaMA-7B | average gap vs FT: -9.2 -> -9.2 | No clear effect |
| use TeZO instead of MeZO | TeZO · RoBERTa-large k=16 | average gap vs FT: -8.7 -> -9.3 | Hurt |
| use TeZO instead of MeZO | TeZO · RoBERTa-large k=512 | average gap vs FT: -3.3 -> -3.6 | Hurt |
| use TeZO instead of MeZO | TeZO · OPT-13B | average gap vs FT: -5.4 -> -5.2 | Improved |
| use TeZO-Adam instead of MeZO-Adam | TeZO-Adam · OPT-13B RTE memory | GPU memory: 75.27 G -> 26.01 G | Improved |
| use TeZO-Adam instead of MeZO-Adam | TeZO-Adam · OPT-13B RTE wall-clock | time per iteration: 642 ms -> 394 ms | Improved |
| use TeZO instead of MeZO on small models | TeZO · OPT-125M wall-clock | time per iteration: 33 ms -> 36 ms | Hurt |
| use TeZO instead of MeZO on large models | TeZO · OPT-30B wall-clock | time per iteration: 871 ms -> 680 ms | Improved |
Closest papers by meaning
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · similarity 0.93)
- Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning (NeurIPS 2025 · similarity 0.93)
- 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 0.91)
- Towards Efficient Low-Order Hybrid Optimizer for Language Model Fine-Tuning (AAAI 2025 · similarity 0.90)
- Variance-reduced Zeroth-Order Methods for Fine-Tuning Language Models (ICML 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.
