Research map / Architectures and efficiency / Pruning and efficient tuning
Less Is More: Rethinking Parameter-Efficient Fine-Tuning from a Subtractive Perspective
Tianqi Jiang, Liu Yang, Xi-Le Zhao, Zixuan Qin, Qinghua Hu
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 |
|---|---|---|---|
| replace a layer with a pseudo-inverse-initialized filter block instead of training that layer directly (ablation) | ViT-B/16 on VTAB-1k · VTAB-1k single-layer | average accuracy (%): 68.35 -> 73.65 (params 7.09M -> 0.63M) | Improved |
| tune LayerNorm parameters (ablation) | SFP · VTAB-1k Natural (single-layer) | accuracy (%): 76.90 -> 81.24 | Improved |
| tune LayerNorm parameters (ablation) | SFP · VTAB-1k Specialized (single-layer) | accuracy (%): 86.44 -> 86.65 | Improved |
| tune LayerNorm parameters (ablation) | SFP · VTAB-1k Structured (single-layer) | accuracy (%): 59.59 -> 60.51 | Improved |
| use pseudo-inverse initialization instead of Kaiming init (ablation) | SFP filter block · VTAB-1k Natural (single-layer) | accuracy (%): 80.01 -> 81.24 | Improved |
| use pseudo-inverse initialization instead of Kaiming init (ablation) | SFP filter block · VTAB-1k Specialized (single-layer) | accuracy (%): 85.85 -> 86.65 | Improved |
| use pseudo-inverse initialization instead of Kaiming init (ablation) | SFP filter block · VTAB-1k Structured (single-layer) | accuracy (%): 59.49 -> 60.51 | Improved |
| substitute two layers instead of one (ablation) | SFP · VTAB-1k | average accuracy (%): 73.65 -> 73.03 | Hurt |
| use SFP instead of full fine-tuning | ViT-B/16 · VTAB-1k | mean accuracy (%): 65.57 -> 74.04 | Improved |
| use SFP instead of DTL (best PEFT baseline) | ViT-B/16 · VTAB-1k | mean accuracy (%): 73.26 -> 74.04 | Improved |
| use SFP instead of GPS on datasets close to pretraining data | ViT-B/16 · CIFAR-100 and SUN397 (VTAB-1k) | accuracy (%): CIFAR-100 74.0 -> 70.5; SUN397 56.6 -> 51.9 | Hurt |
| replace a layer with filter block (inference efficiency) | ViT-B/16 · VTAB-1k | inference time per image (ms): 2.81 -> 2.33 | Improved |
Closest papers by meaning
- Faster Parameter-Efficient Tuning with Token Redundancy Reduction (CVPR 2025 · similarity 0.88)
- Parameter Efficient Fine-tuning via Explained Variance Adaptation (NeurIPS 2025 · similarity 0.88)
- SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors (NeurIPS 2024 · similarity 0.88)
- From PEFT to DEFT: Parameter Efficient Finetuning for Reducing Activation Density in Transformers (AAAI 2025 · similarity 0.88)
- PACE: Marrying generalization in PArameter-efficient fine-tuning with Consistency rEgularization (NeurIPS 2024 · similarity 0.88)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.87)
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.
