Research map / Architectures and efficiency / Pruning and efficient tuning
Stable-LoRA: Stabilizing Feature Learning of Low-Rank Adaption
Yize Wu, KE GAO, Ling Li, Yanjun Wu
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 weight shrinkage of A (Stable-LoRA) | LoRA with AdamW · QA average, 4 models | accuracy: 61.94 -> 64.01 (Qwen2-0.5B) | Improved |
| add weight shrinkage of A (Stable-LoRA) | LoRA with AdamW · GSM8K, Llama-3.2-3B, 5000 steps | accuracy: 58.83 -> 59.44 | Improved |
| use LoRA+ (larger learning rate for B) | LoRA with AdamW · QA average | accuracy: 61.94 -> 62.55 (0.5B); 71.90 -> 71.36 (1B) | Mixed |
| use Riemann preconditioning | LoRA with AdamW · QA average | accuracy: 71.90 -> 69.07 (Llama-3.2-1B) | Hurt |
| apply Stable-LoRA to more target modules (qkvogud) (ablation) | LoRA with AdamW · Combined QA | accuracy: 64.35 -> 64.75 (0.5B) | Improved |
| stop shrinking at the stable condition (ablation) | Stable-LoRA · Combined QA | accuracy: 62.46 vs 62.52 (0.5B) | No clear effect |
| use a large shrinkage ratio lambda=0.005 (ablation) | Stable-LoRA · Combined QA (qv) | accuracy: 72.29 vs 72.80 (1B, lambda 0.0005) | Mixed |
| add Stable-LoRA to Llama-3.1-8B | LoRA with AdamW · Math reasoning, Llama-3.1-8B qkvogud | accuracy: 71.72 -> 74.37 (3000 steps) | Improved |
| use Stable-LoRA instead of DoRA | LoRA · QA average, Llama-3.2-1B | accuracy: 71.68 -> 72.52 | Improved |
| add Stable-LoRA shrinkage | LoRA with AdamW · HellaSwag, Qwen2-0.5B | training time: 217.4 -> 218.8 | No clear effect |
Closest papers by meaning
- LoRA+: Efficient Low Rank Adaptation of Large Models (ICML 2024 · similarity 0.92)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.91)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.91)
- The Expressive Power of Low-Rank Adaptation (ICLR 2024 · similarity 0.91)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.91)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.90)
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.
