Research map / Architectures and efficiency / Pruning and efficient tuning
Sparse Fine-Tuning of Transformers for Generative Tasks
Wei Chen, Jingxi Yu, Zichen Miao, Qiang Qiu
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 |
|---|---|---|---|
| adapt atoms instead of coefficients in toy signal reconstruction (ablation) | attention block · toy high-frequency signal task | qualitative reconstruction | Improved |
| tune only coefficients with atoms fixed (ablation) | attention block · toy high-frequency signal task | qualitative reconstruction | Hurt |
| add sparse coefficients to attention output (ablation) | attention block · toy signal reconstruction | atoms needed: 100 -> 9 atoms | Improved |
| use sparse dictionary fine-tuning instead of LoRA | PixArt-Sigma · DreamBooth customization | text alignment: 0.218 -> 0.235 | Improved |
| use sparse dictionary fine-tuning instead of LoRA | PixArt-Sigma · DreamBooth customization | diversity (Vendi): 4.707 -> 5.205 | Improved |
| use rho=0.02 instead of LoRA for ImageReward | PixArt-Sigma · DreamBooth customization | ImageReward: 0.220 -> 0.182 | Hurt |
| increase coefficient density rho from 0.02 to 0.04 (ablation) | sparse fine-tuning · DreamBooth customization | ImageReward: 0.182 -> 0.431 | Improved |
| increase dictionary size M from 16 to 2048 (ablation) | sparse fine-tuning · DreamBooth customization, rho=0.01 | fidelity: 0.547 -> 0.692 | Improved |
| increase dictionary size M from 16 to 2048 (ablation) | sparse fine-tuning · DreamBooth customization, rho=0.01 | diversity (Vendi): 5.867 -> 3.525 | Hurt |
| increase sparsity (lower density rho) (ablation) | sparse fine-tuning · DreamBooth customization | fidelity | Mixed |
| remove unimportant atoms by raising sparsity at edit time | fine-tuned PixArt-Sigma · image editing | qualitative text alignment of edit | Improved |
| sparse atom adaptation versus rank-1 LoRA under modified prompt (ablation) | PixArt-Sigma · image editing | stability | 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.
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 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
- WeGeFT: Weight‑Generative Fine‑Tuning for Multi‑Faceted Efficient Adaptation of Large Models (ICML 2025 · similarity 0.85)
- In-Context Compositional Learning vis Sparse Coding Transformer (NeurIPS 2025 · similarity 0.84)
- Sparse is Enough in Fine-tuning Pre-trained Large Language Models (ICML 2024 · similarity 0.84)
- Learning Vision and Language Concepts for Controllable Image Generation (ICML 2025 · similarity 0.84)
- Large Convolutional Model Tuning via Filter Subspace (ICLR 2025 · similarity 0.83)
- S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity (NeurIPS 2024 · similarity 0.83)
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.
