Research map / Architectures and efficiency / Pruning and efficient tuning
Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation
Nairouz Mrabah, Nicolas Richet, Ismail Ben Ayed, Eric Granger
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 |
|---|---|---|---|
| sparsify the gradient (random Top-M selection) instead of dense Adam (ablation) | SO · 9 datasets 1-shot | average top-1 accuracy: 4.4 -> 73.6 | Improved |
| add importance-pruned sparse moments (ablation) | SO · 9 datasets 1-shot | average top-1 accuracy: 73.6 -> 74.6 | Improved |
| add importance-pruned sparse moments (ablation) | SO · 9 datasets 4-shot | average top-1 accuracy: 77.2 -> 80.4 | Improved |
| use random instead of importance-based gradient selection (ablation) | SO · 9 datasets 1-shot | average top-1 accuracy: 66.6 -> 74.6 | Improved |
| use random instead of importance-based gradient selection (ablation) | SO · 9 datasets 2-shot | average top-1 accuracy: 68.3 -> 77.7 | Improved |
| use importance-based instead of random moment pruning (ablation) | SO · 9 datasets 1-shot | average top-1 accuracy: 72.9 -> 74.6 | Improved |
| use importance-based instead of random moment pruning (ablation) | SO · 9 datasets 4-shot | average top-1 accuracy: 79.8 -> 80.4 | Improved |
| use dynamic instead of static sparsity support (ablation) | SO · 9 datasets 1-shot | average top-1 accuracy: 71.2 -> 74.6 | Improved |
| use dynamic instead of static sparsity support (ablation) | SO · 9 datasets 4-shot | average top-1 accuracy: 77.0 -> 80.4 | Improved |
| add moments (ablation) | SO · 9 datasets 1-shot | average top-1 accuracy: 67.8 -> 74.6 | Improved |
| add moments (ablation) | SO · 9 datasets 2-shot | average top-1 accuracy: 73.0 -> 77.7 | Improved |
| use SO instead of ReLoRA | SO · 11 datasets 1-shot | average top-1 accuracy: 72.5 -> 73.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 5 of 5 on most shared results.
- ReLoRA — ReLoRA: High-Rank Training Through Low-Rank Updates (ICLR 2024) · better on 3 of 3 shared results
- VeRA — VeRA: Vector-based Random Matrix Adaptation (ICLR 2024) · better on 1 of 1 shared result
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 1 of 1 shared result
- GaLoRE — GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (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
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.90)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.90)
- SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025 · similarity 0.89)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.89)
- CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models (ICML 2025 · similarity 0.89)
- ECoFLaP: Efficient Coarse-to-Fine Layer-Wise Pruning for Vision-Language Models (ICLR 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.
