Research map / Architectures and efficiency / Pruning and efficient tuning
Task-Specific Directions: Definition, Exploration, and Utilization in Parameter Efficient Fine-Tuning
Chongjie Si, Z. Shi, Shifan Zhang, Xiaokang Yang, Hanspeter Pfister, Wang 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 |
|---|---|---|---|
| add LoRA-Dash | LoRA r=32 · Commonsense (LLaMA3-8B) | avg accuracy: 80.8 -> 85.4 | Improved |
| add LoRA-Init | LoRA r=32 · Commonsense (LLaMA3-8B) | avg accuracy: 80.8 -> 83.4 | Improved |
| add LoRA-Dash and LoRA-Init (LoRA-TSD) | LoRA r=32 · Commonsense (LLaMA3-8B) | avg accuracy: 80.8 -> 85.5 | Improved |
| add LoRA-TSD | LoRA r=4 · Commonsense (LLaMA-7B) | avg accuracy: 55.1 -> 76.1 | Improved |
| add LoRA-TSD | LoRA r=64 · Commonsense (LLaMA-7B) | avg accuracy: 65.8 -> 77.0 | Improved |
| add LoRA-Dash | AdaLoRA r=32 · Commonsense (LLaMA-7B) | avg accuracy: 75.5 -> 76.0 | Improved |
| add LoRA-Dash | FLoRA r=32 · Commonsense (LLaMA-7B) | avg accuracy: 74.3 -> 76.7 | Improved |
| add LoRA-TSD | LoRA r=2 · GLUE (DeBERTaV3-base) | avg score: 88.13 -> 89.88 | Improved |
| add LoRA-TSD | LoRA r=2 · GLUE (DeBERTaV3-large) | avg score: 90.41 -> 91.35 | Improved |
| initialize LoRA with TSDs instead of bottom singular vectors (MiLoRA) (ablation) | LoRA · Commonsense (LLaMA3-8B) | avg accuracy: 81.9 -> 83.4 | Improved |
| initialize LoRA with TSDs instead of top singular vectors (PiSSA) (ablation) | LoRA · Commonsense (LLaMA3-8B) | avg accuracy: 75.4 -> 83.4 | Improved |
| dash on TSDs instead of top, random or bottom directions (ablation) | LoRA-Dash · Commonsense (LLaMA3-8B) | accuracy | Improved |
Closest papers by meaning
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · similarity 0.90)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.90)
- LoRA vs Full Fine-tuning: An Illusion of Equivalence (NeurIPS 2025 · similarity 0.89)
- LoRA-One: One-Step Full Gradient Could Suffice for Fine-Tuning Large Language Models, Provably and Efficiently (ICML 2025 · similarity 0.89)
- HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning (NeurIPS 2024 · similarity 0.89)
- VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks (NeurIPS 2024 · 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.
