Research map / Architectures and efficiency / Pruning and efficient tuning
Generalized Tensor-based Parameter-Efficient Fine-Tuning via Lie Group Transformations
Chongjie Si, Zhiyi Shi, Xuehui Wang, Yichen Xiao, Xiaokang Yang, Wei Shen
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 additive LoRA update with Lie group multiplicative update (LieRA) | LoRA r=16 ConvNeXt-V2-B · VTAB-1k average | accuracy: 74.1 -> 75.5 | Improved |
| replace additive LoRA update with LieRA | LoRA r=2 ConvNeXt-V2-B · VTAB-1k average | accuracy: 74.4 -> 74.7 | Improved |
| replace additive LoRA update with LieRA | LoRA r=16 ConvNeXt-V2-B · MS COCO detection | mAP: 35.5 -> 39.1 | Improved |
| replace additive LoRA update with LieRA | LoRA r=32 ConvNeXt-V2-B · MS COCO detection | mAP: 35.9 -> 40.5 | Improved |
| replace additive LoRA update with LieRA | LoRA r=16 LLaMA-7B · Commonsense reasoning average | accuracy: 70.9 -> 75.2 | Improved |
| replace additive LoRA update with LieRA | LoRA r=32 LLaMA3-8B · Commonsense reasoning average | accuracy: 80.8 -> 85.3 | Improved |
| replace additive LoRA update with LieRA | LoRA r=2 DeBERTaV3-base · GLUE dev average | average score: 88.13 -> 88.97 | Improved |
| replace additive LoRA update with LieRA | LoRA r=8 DeBERTaV3-large · GLUE dev average | average score: 90.79 -> 90.98 | Mixed |
| replace additive LoRA update with LieRA on LLaMA-7B | LoRA r=16 LLaMA-7B · BoolQ | accuracy: 69.9 -> 68.5 | Hurt |
| add first-order Taylor approximation to LieRA (ablation) | LieRA r=16 · VTAB-1k average | accuracy: 75.7 -> 75.5 | No clear effect |
| add first-order Taylor approximation to LieRA (ablation) | LieRA r=16 · MS COCO training cost | time per epoch (min): 76.28 -> 50.17 | Improved |
| add first-order Taylor approximation to LieRA (ablation) | LieRA r=16 · MS COCO training cost | GPU memory (GB): 14.74 -> 9.97 | 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 (0.70%) — PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024) · better on 1 of 1 shared result
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
- SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors (NeurIPS 2024 · similarity 0.90)
- Maintaining Structural Integrity in Parameter Spaces for Parameter Efficient Fine-tuning (ICLR 2025 · similarity 0.88)
- Structured Unrestricted-Rank Matrices for Parameter Efficient Finetuning (NeurIPS 2024 · similarity 0.88)
- GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning (NeurIPS 2025 · similarity 0.87)
- Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation (CVPR 2025 · similarity 0.87)
- A Stronger Mixture of Low-Rank Experts for Fine-Tuning Foundation Models (ICML 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.
