Research map / Architectures and efficiency / Pruning and efficient tuning
Three Forward, One Backward: Memory-Efficient Full-Rank Fine-Tuning of Large Models via Extra Forward Passes
Jia Zhang, Yu Bai, Hualin Zhang, Tianshuo Chen, Zhaogeng Liu, zhiqiang xu, Yi Chang, Bin Gu
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 zeroth-order full-weight updates to LoRA (alternating) | LoRA · SST-2 (OPT-1.3B) | accuracy: 92.2 -> 92.4 | Improved |
| add zeroth-order full-weight updates to LoRA (alternating) | LoRA · BoolQ (OPT-6.7B) | accuracy: 49.4 -> 65.1 | Improved |
| add zeroth-order full-weight updates to LoRA | LoRA · RTE (OPT-1.3B) | accuracy: 61.1 -> 59.4 | Hurt |
| add LoRA updates to MeZO | MeZO · SST-2 (OPT-1.3B) | accuracy: 58.1 -> 92.4 | Improved |
| add LoRA updates to MeZO | MeZO · SST-2 (RoBERTa-large k=16) | accuracy: 90.5 -> 92.8 | Improved |
| alternate low-rank and ZO updates (full LMAO) vs components alone (w/o LoRA or w/o MeZO) (ablation) | LMAO · OPT-1.3B six tasks | accuracy | Improved |
| increase training data from k=16 to k=512 per class | LMAO · RoBERTa-large (6 tasks) | accuracy: gap to FT narrows; SNLI 87.9 vs FT 88.7 | Mixed |
| vary LoRA rank r and scaling alpha (ablation) | LMAO · SST-2 (RoBERTa-large) | accuracy | No clear effect |
| add LoRA update on top of ZO step (alternating) memory | LMAO · peak GPU memory BoolQ OPT-1.3B | GB: 38.48 -> 23.08 | Improved |
| use LMAO versus LoRA/MeZO on OPT-6.7B WIC | LMAO · WIC (OPT-6.7B) | accuracy: 52.9 -> 50.0 | Hurt |
Closest papers by meaning
- Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures (ICLR 2025 · similarity 0.94)
- Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.93)
- GoRA: Gradient-driven Adaptive Low Rank Adaptation (NeurIPS 2025 · similarity 0.93)
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · similarity 0.93)
- AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning (ICLR 2025 · similarity 0.92)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.92)
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.
