Research map / Large language models / Instruction and fine-tuning
SG-LoRA: Semantic-guided LoRA Parameters Generation
Miaoge Li, Yang Chen, Zhijie Rao, Can Jiang, Kang Wei, Jingcai 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 |
|---|---|---|---|
| generate LoRA with SG-LoRA instead of Top-k weighted merging | Top-k LoRA Weighted · MS-COCO retrieval | I2T R@1: 71.55 -> 74.31 | Improved |
| generate LoRA with SG-LoRA instead of Oracle fine-tuned LoRA | SG-LoRA · MS-COCO retrieval | T2I R@1: 53.10 -> 54.42 | Improved |
| generate LoRA with SG-LoRA instead of Oracle fine-tuned LoRA | SG-LoRA · OxfordPets retrieval | I2T R@1: 55.84 -> 57.15 | Improved |
| weight top-k experts with softmax similarity instead of equal weights (ablation) | Top-k LoRA merging · MS-COCO retrieval | I2T R@1: 70.70 -> 71.55 | Improved |
| select semantic top-k experts instead of uniformly averaging all (Model Soups) (ablation) | Model Soups · MS-COCO retrieval | T2I R@1: 47.38 -> 49.85 | Improved |
| train the generator on MS-COCO and apply to OxfordPets | SG-LoRA · OxfordPets retrieval | I2T R@1: 57.15 -> 55.41 | Hurt |
| train generator on OxfordPets and test on MS-COCO | SG-LoRA · MS-COCO retrieval | I2T R@1: 74.31 -> 70.81 | Hurt |
| use textual semantic prior instead of visual prior (ablation) | SG-LoRA · MS-COCO retrieval | I2T R@1: 73.16 -> 74.31 | Improved |
| use textual semantic prior instead of visual prior (ablation) | SG-LoRA · CIFAR-100 superclass | accuracy: 73.83 -> 77.50 | Improved |
| use detailed description listing classes instead of generic photo prompt (ablation) | SG-LoRA · CIFAR-100 superclass | accuracy: 75.77 -> 77.50 | Improved |
| add MS-COCO Cat expert to the repository (ablation) | SG-LoRA · OxfordPets Egyptian Mau | I2T R@1: 36.08 -> 37.11 | Improved |
| add MS-COCO Cat expert to the repository (ablation) | SG-LoRA · OxfordPets Persian | I2T R@5: 80.00 -> 79.65 | Hurt |
Closest papers by meaning
- Merging LoRAs like Playing LEGO: Pushing the Modularity of LoRA to Extremes Through Rank-Wise Clustering (ICLR 2025 · similarity 0.90)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.90)
- $\textit{Trans-LoRA}$: towards data-free Transferable Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.90)
- LoRA-Gen: Specializing Large Language Model via Online LoRA Generation (ICML 2025 · similarity 0.89)
- BSLoRA: Enhancing the Parameter Efficiency of LoRA with Intra-Layer and Inter-Layer Sharing (ICML 2025 · similarity 0.89)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.89)
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.
