Research map / Architectures and efficiency / Pruning and efficient tuning
Beyond Low-Rank Tuning: Model Prior-Guided Rank Allocation for Effective Transfer in Low-Data and Large-Gap Regimes.
Chuyan Zhang, Kefan Wang, Yun 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 |
|---|---|---|---|
| allocate rank per layer by stable rank instead of uniform rank 256 | LoRA · VTAB-Specialized 1-shot | accuracy (%): 54.35 -> 56.38 | Improved |
| allocate rank by stable rank instead of uniform rank 64 (similar trainable parameters) (ablation) | LoRA · VTAB-Specialized 1-shot | accuracy (%): 48.31 -> 57.69 | Improved |
| raise LoRA rank from 8 to 256 | LoRA · VTAB-Specialized 1-shot | accuracy (%): 42.99 -> 54.35 | Improved |
| add stochastic partial updating (SPU) to fixed-rank LoRA r256 (ablation) | LoRA-r256 · VTAB-Specialized 1-shot | accuracy (%): 54.35 -> 53.69 | No clear effect |
| add stochastic partial updating to fixed rank 8 (ablation) | LoRA-r8 · MedFM 10-shot | AUC (%): 74.84 -> 78.82 | Improved |
| add stochastic partial updating to SR-LoRA (ablation) | SR-LoRA · MedFM 1-shot | AUC (%): 71.32 -> 72.47 | Improved |
| add stochastic partial updating to SR-LoRA (ablation) | SR-LoRA · VTAB-Specialized 1-shot | accuracy (%): 57.69 -> 56.38 | Hurt |
| replace full fine-tuning with SR-LoRA on ViT-L | ViT-L (ImageNet-21k) · VTAB-Specialized 1-shot | accuracy (%): 46.26 -> 53.37 | Improved |
| replace LoRA with SR-LoRA on DINO ViT-L | DINO ViT-L · VTAB-Specialized 1-shot | accuracy (%): 36.44 -> 43.59 | Improved |
| replace LoRA with SR-LoRA on MAE ViT-L | MAE ViT-L · VTAB-Specialized 1-shot | accuracy (%): 32.80 -> 46.10 | Improved |
| replace full fine-tuning with SR-LoRA | ViT-B/16 · MedFM 1-shot | AUC (%): 67.31 -> 72.47 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Pissa — PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024) · better on 4 of 4 shared results
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
- LoRA-Pro: Are Low-Rank Adapters Properly Optimized? (ICLR 2025 · similarity 0.93)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.93)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.93)
- The Expressive Power of Low-Rank Adaptation (ICLR 2024 · similarity 0.93)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.92)
- Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation (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.
