Research map / Architectures and efficiency / Pruning and efficient tuning
Fine-Tuning Language Models with Just Forward Passes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| MeZO fine-tuning instead of zero-shot | OPT-13B · SST-2 | accuracy: 58.8 -> 91.4 | Helped |
| MeZO instead of in-context learning | OPT-13B · SQuAD | F1: 75.9 -> 84.7 | Helped |
| MeZO instead of backprop fine-tuning (12x memory) | OPT-13B · SQuAD | F1: 84.9 -> 84.7 | No clear effect |
| MeZO (prefix) instead of FT | OPT-13B · CB | accuracy: 83.9 -> 69.6 | Hurt |
| MeZO instead of FT | OPT-13B · COPA | accuracy: 79.0 -> 88.0 | Helped |
| MeZO scaled to 66B | OPT-66B · SST-2 | accuracy: 89.3 -> 93.6 | Helped |
| use a prompt for MeZO fine-tuning (ablation) | RoBERTa-large · SST-2 | accuracy: 51.9 (2.9) -> 89.6 (1.2) | Helped |
| use a prompt for MeZO fine-tuning (ablation) | RoBERTa-large · TREC | accuracy: 19.5 (9.0) -> 66.7 (6.2) | Helped |
| optimize accuracy objective instead of cross entropy with MeZO (ablation) | RoBERTa-large · TREC | accuracy: 94.3 -> 68.6 | Hurt |
| optimize accuracy instead of cross entropy with MeZO (ablation) | RoBERTa-large · SST-2 | accuracy: 93.3 -> 92.7 | No clear effect |
| increase SPSA samples n from 1 to 16 at fixed forward-pass budget (ablation) | MeZO · SST-2/SNLI/TREC | accuracy: 65.1 -> 67.0 (SNLI); 66.7 -> 62.8 (TREC) | No clear effect |
| MeZO with LoRA/prefix instead of full-parameter (ablation) | MeZO · OPT-13B tasks | accuracy: 91.4 -> 89.6 (SST-2) | No clear effect |
Closest papers by meaning
- Variance-reduced Zeroth-Order Methods for Fine-Tuning Language Models (ICML 2024 · similarity 0.96)
- Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning (NeurIPS 2025 · similarity 0.94)
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · similarity 0.93)
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 2024 · similarity 0.92)
- Towards Efficient Low-Order Hybrid Optimizer for Language Model Fine-Tuning (AAAI 2025 · similarity 0.92)
- Harmony in Divergence: Towards Fast, Accurate, and Memory-efficient Zeroth-order LLM Fine-tuning (NeurIPS 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.