Research map / Architectures and efficiency / Pruning and efficient tuning
Gradient Intrinsic Dimensionality Alignment:Narrowing The Gap Between Low-Rank Adaptation and Full Fine-Tuning
Jingqi Ye, haonan he, Minglei Li, Fujun Han, Tao Chen, Peng Ye
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 |
|---|---|---|---|
| align LoRA rank with gradient intrinsic dimensionality and reallocate across layers (RaLoRA-Pro) | LoRA · GLUE average (T5-Base) | accuracy: 82.08 -> 87.23 | Improved |
| align LoRA rank with gradient intrinsic dimensionality (RaLoRA) | LoRA · GSM8K (LLaMA-3.1-8B) | accuracy: 67.78 -> 72.25 | Improved |
| align LoRA rank with gradient intrinsic dimensionality (RaLoRA) | LoRA · HumanEval (LLaMA-3.1-8B) | pass@1: 43.09 -> 48.78 | Improved |
| add inter-layer reallocation on top of alignment (RaLoRA-Pro) | LoRA · MT-Bench (LLaMA-3.1-8B) | score: 6.15 -> 6.72 | Improved |
| RaLoRA alignment on CLIP image classification | LoRA · Image classification average (CLIP-ViT-B/16) | accuracy: 89.08 -> 90.53 | Improved |
| use entropy-based estimator instead of SVD threshold 1e-3 (ablation) | RaLoRA · GSM8K | accuracy: 70.64 -> 72.25 | Improved |
| use entropy-based estimator instead of SVD threshold 1e-1 (ablation) | RaLoRA · HumanEval | pass@1: 45.53 -> 48.78 | Improved |
| add intra-layer GID alignment to loss-sensitivity reallocation (LS-LoRA to RaLoRA-Pro, rank 8) (ablation) | LS-LoRA · HumanEval | pass@1: 39.83 -> 46.95 | Improved |
| add GID alignment to LoRA at rank 64 (ablation) | LoRA · GSM8K | accuracy: 67.17 -> 74.55 | Improved |
| add inter-layer reallocation to RaLoRA at rank 64 (ablation) | RaLoRA · GSM8K | accuracy: 74.55 -> 75.23 | Improved |
| increase max expansion factor nmax from 16 to 64 (ablation) | RaLoRA · GSM8K / HumanEval | accuracy / pass@1: GSM8K 72.25 -> 71.19; HumanEval 45.12 -> 48.78 | Mixed |
| widen rank reallocation range from 8-8 to 4-32 (ablation) | RaLoRA-Pro · MT-Bench / GSM8K | score / accuracy: MT-Bench 6.19 -> 6.70; GSM8K 71.42 -> 71.39 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 1 of 1 shared result
- PiSSA — 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
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.92)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.91)
- GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning (NeurIPS 2025 · similarity 0.90)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.90)
- LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · similarity 0.89)
- LoRA-Pro: Are Low-Rank Adapters Properly Optimized? (ICLR 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.
