Research map / Architectures and efficiency / Pruning and efficient tuning
AIRA: Activation-Informed Low-Rank Adaptation for Large Models
Lujun Li, Dezhi Li, Cheng Lin, Wei Li, Wei Xue, Sirui Han, Yike Guo
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 outlier-weighted SVD initialization instead of Kaiming (ablation) | AIRA · few-shot 4-shot | top-1 accuracy: 79.8 -> 80.2 | Improved |
| add outlier-weighted SVD initialization instead of Kaiming (ablation) | AIRA · few-shot 16-shot | top-1 accuracy: 83.6 -> 84.6 | Improved |
| add outlier-weighted SVD initialization instead of standard SVD (ablation) | AIRA · few-shot 4-shot | top-1 accuracy: 79.9 -> 80.2 | Improved |
| add outlier-weighted SVD initialization instead of standard SVD (ablation) | AIRA · few-shot 16-shot | top-1 accuracy: 84.3 -> 84.6 | Improved |
| add outlier-driven rank assignment instead of uniform (ablation) | AIRA · few-shot 4-shot | top-1 accuracy: 78.2 -> 80.2 | Improved |
| add outlier-driven rank assignment instead of uniform (ablation) | AIRA · few-shot 16-shot | top-1 accuracy: 83.8 -> 84.6 | Improved |
| add outlier-driven rank assignment instead of growth (ablation) | AIRA · few-shot 16-shot | top-1 accuracy: 84.0 -> 84.6 | Improved |
| add input-weighted normalization instead of original LoRA training (ablation) | AIRA · few-shot 4-shot | top-1 accuracy: 79.0 -> 80.2 | Improved |
| add input-weighted normalization instead of output-weighted (ablation) | AIRA · few-shot 16-shot | top-1 accuracy: 84.0 -> 84.6 | Improved |
| use -x objective instead of -log(x) (ablation) | AIRA · few-shot 4-shot | top-1 accuracy: 65.5 -> 64.4 | Hurt |
| use learning rate 1e-3 instead of 1e-4 (ablation) | AIRA · few-shot 16-shot | top-1 accuracy: 70.8 -> 73.4 | Improved |
| use AIRA instead of PiSSA on Mistral-7B | AIRA · GSM8K+MATH | top-1 accuracy: 44.35 -> 45.23 | Improved |
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.
- PiSSA — PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models (NeurIPS 2024) · better on 9 of 10 shared results
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 2 of 2 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
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.91)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.91)
- LoRA+: Efficient Low Rank Adaptation of Large Models (ICML 2024 · similarity 0.90)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.90)
- LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · similarity 0.89)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (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.
