Research map / Architectures and efficiency / Pruning and efficient tuning
Efficient Fine-Tuning of Large Models via Nested Low-Rank Adaptation
Lujun Li, Cheng Lin, Dezhi Li, You-Liang Huang, Wei Li, Tianyu Wu, Jie Zou, Wei Xue, Sirui Han, Yike Guo
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 AwSVD for outer LoRA init instead of random (ablation) | NoRA · GSM8K | accuracy: 66.4 -> 69.4 | Improved |
| use AwSVD instead of plain SVD for outer init (ablation) | NoRA · GSM8K | accuracy: 69.1 -> 69.4 | Improved |
| use AwSVD instead of plain SVD for outer init (ablation) | NoRA · MATH | accuracy: 18.9 -> 19.1 | Improved |
| use Gaussian inner init instead of uniform/zero (ablation) | NoRA · GSM8K | accuracy: 68.3 -> 69.4 | Improved |
| use Gaussian inner init instead of diag(Sigma) (ablation) | NoRA · GSM8K | accuracy: 68.9 -> 69.4 | Improved |
| use serial LoRA instead of parallel LoRA (ablation) | NoRA · GSM8K | accuracy: 66.4 -> 69.4 | Improved |
| use serial LoRA instead of adapter (ablation) | NoRA · GSM8K | accuracy: 68.0 -> 69.4 | Improved |
| place inner LoRA at wa only instead of wa and wb (ablation) | NoRA · GSM8K | accuracy: 68.9 -> 69.4 | Improved |
| place inner LoRA at wa instead of wb (ablation) | NoRA · GSM8K | accuracy: 68.4 -> 69.4 | Improved |
| increase inner rank r' from 8 to 64 at outer rank 64 (ablation) | NoRA · MATH | accuracy: 17.10 -> 19.48 | Improved |
| increase inner rank r' from 8 to 64 at outer rank 128 (ablation) | NoRA · MATH | accuracy: 17.80 -> 19.83 | Improved |
| use NoRA instead of LoRA-XS (r=128) | LLaVA-1.5-7B · TextVQA | accuracy: 54.36 -> 58.70 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- VeRA — VeRA: Vector-based Random Matrix Adaptation (ICLR 2024) · better on 2 of 2 shared results
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.92)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.92)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.92)
- LoRA-Pro: Are Low-Rank Adapters Properly Optimized? (ICLR 2025 · similarity 0.91)
- Beyond Zero Initialization: Investigating the Impact of Non-Zero Initialization on LoRA Fine-Tuning Dynamics (ICML 2025 · similarity 0.91)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · 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.
