Research map / Architectures and efficiency / Pruning and efficient tuning
Put the Space of LoRA Initialization to the Extreme to Preserve Pre-trained Knowledge
Pengwei Tang, Xiaolin Hu, Yong Liu, Lizhong Ding, Dongjie Zhang, Xing Wu, Debing Zhang
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 |
|---|---|---|---|
| initialize LoRA in the null space of activations instead of CorDA initialization | LoRA · Math, LLaMA-2-7B | Avg1(Per): 76.24 -> 79.21 | Improved |
| initialize LoRA in the null space of activations instead of CorDA initialization | LoRA · Code, LLaMA-2-7B | Avg1(Per): 91.42 -> 95.12 | Improved |
| initialize LoRA in the null space of activations instead of CorDA initialization | LoRA · Instruction following, LLaMA-2-7B | Avg1(Per): 87.36 -> 89.90 | Improved |
| initialize LoRA in the null space of activations instead of CorDA initialization | LoRA · Math, LLaMA-3.2-3B | Avg1(Per): 89.00 -> 93.18 | Improved |
| initialize LoRA in the null space of activations instead of CorDA initialization | LoRA · Math, LLaMA-3.2-3B | Avg2 (downstream): 36.81 -> 36.41 | Hurt |
| use LoRA-Null initialization instead of PiSSA | LoRA · Math, LLaMA-2-7B | Avg2 (downstream): 29.67 -> 26.62 | Hurt |
| reduce calibration set to 64 samples (CorDA) (ablation) | CorDA · Math, LLaMA-3.2-3B | Avg1(Per): 89.00 -> 68.97 | Hurt |
| reduce calibration set to 64 samples (LoRA-Null) (ablation) | LoRA-Null · Math, LLaMA-3.2-3B | Avg1(Per): 93.00 -> 94.95 | No clear effect |
| increase LoRA rank from 128 to 256 (CorDA) (ablation) | CorDA · Math, LLaMA-3.2-3B | Avg1(Per): 89.00 -> 73.01 | Hurt |
| increase LoRA rank from 128 to 256 (LoRA-Null) (ablation) | LoRA-Null · Math, LLaMA-3.2-3B | Avg1(Per): 93.00 -> 84.71 | Hurt |
| scale singular values of LoRA-Null init by 4 (ablation) | LoRA-Null · Math, Gemma-3-1B | Avg1(Per): 80.78 -> 65.98 | Hurt |
| scale singular values of LoRA-Null init by 4 (ablation) | LoRA-Null · Math, Gemma-3-1B | GSM8k: 16.91 -> 22.90 | 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 — PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024) · better on 4 of 6 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
- LoRA vs Full Fine-tuning: An Illusion of Equivalence (NeurIPS 2025 · similarity 0.94)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.92)
- LoRA Training in the NTK Regime has No Spurious Local Minima (ICML 2024 · similarity 0.92)
- Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning (NeurIPS 2026 · similarity 0.91)
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · similarity 0.91)
- Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation (NeurIPS 2025 · similarity 0.91)
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.
