Research map / Architectures and efficiency / Pruning and efficient tuning
DWTSG: Parameter-Efficient Fine-Tuning of Large Pre-trained Models via Discrete Wavelet Transform and Subband Guidance
Chengwei Sun, Jiwei Wei, Shiyuan He, Zeyu Ma, Yuyang Zhou, Ran Ran, Jie Zou, Yang Yang
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 DWTSG instead of LoRA with fewer parameters | RoBERTa-large · GLUE avg | average score: 88.6 -> 88.8 | Improved |
| use DWTSG instead of FourierFT | RoBERTa-large · GLUE avg | average score: 88.2 -> 88.8 | Improved |
| use DWTSG instead of full fine-tuning | RoBERTa-large · GLUE avg | average score: 88.9 -> 88.8 | Hurt |
| use DWTSG instead of MiLoRA | LLaMA2-7B · Commonsense170K avg | average accuracy: 79.2 -> 80.0 | Improved |
| use DWTSG instead of LoRA | ViT-base · image classification avg | accuracy (%): 77.6 -> 82.4 | Improved |
| use DWTSG instead of full fine-tuning | ViT-large · image classification avg | accuracy (%): 90.2 -> 89.6 | Hurt |
| use sym4 wavelet instead of haar, db4 or coif1 (ablation) | DWTSG · ViT-base, four image datasets | accuracy (%): best on all four, up to 5.3 points over other bases | Improved |
| use role-aware subband allocation instead of Random (ablation) | DWTSG · ViT-base OxfordPets | accuracy (%): 95.6 -> 96.7 | Improved |
| use role-aware subband allocation instead of Energy (ablation) | DWTSG · ViT-base FGVC | accuracy (%): 38.5 -> 39.7 | Improved |
| use Energy allocation instead of Random (ablation) | DWTSG · ViT-base OxfordPets | accuracy (%): 95.6 -> 94.6 | Hurt |
| raise scaling factor alpha from 0.3 to 1.1 (ablation) | DWTSG · ViT-base FGVC | accuracy (%): 39.7 -> 5.1 | Hurt |
| raise scaling factor alpha from 0.1 to 0.9 (ablation) | DWTSG · ViT-base OxfordPets and StanfordCars | accuracy (%): 93.1 -> 96.7 (OxfordPets) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- VeRA — VeRA: Vector-based Random Matrix Adaptation (ICLR 2024) · better on 1 of 2 shared results
- PiSSA — PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024) · better on 2 of 2 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
- Parameter-Efficient Fine-Tuning with Discrete Fourier Transform (ICML 2024 · similarity 0.88)
- F-Adapter: Frequency-Adaptive Parameter-Efficient Fine-Tuning in Scientific Machine Learning (NeurIPS 2025 · similarity 0.87)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.87)
- LoCA: Location-Aware Cosine Adaptation for Parameter-Efficient Fine-Tuning (ICLR 2025 · similarity 0.87)
- S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity (NeurIPS 2024 · similarity 0.87)
- Sparse is Enough in Fine-tuning Pre-trained Large Language Models (ICML 2024 · similarity 0.86)
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.
