Research map / Architectures and efficiency / Pruning and efficient tuning
Movement Pruning: Adaptive Sparsity by Fine-Tuning
Victor Sanh, Thomas Wolf, Alexander Rush
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 |
|---|---|---|---|
| replace magnitude pruning with movement pruning | BERT-base · SQuAD dev, 3% weights | F1: 54.5 -> 76.3 | Improved |
| use soft movement pruning (threshold + regularizer) instead of hard Top-v movement pruning (ablation) | BERT-base · SQuAD dev, 3% weights | F1: 76.3 -> 79.9 | Improved |
| replace L0 regularization with movement pruning | BERT-base · SQuAD dev, 3% weights | F1: 73.3 -> 76.3 | Improved |
| use movement pruning instead of magnitude pruning at low sparsity (>70% remaining weights) | BERT-base · MNLI/SQuAD/QQP | accuracy | Hurt |
| add knowledge distillation from dense BERT-base teacher (ablation) | movement pruning · SQuAD dev, 10% weights | F1: 81.7 -> 84.3 | Improved |
| add knowledge distillation (ablation) | soft movement pruning · SQuAD dev, 3% weights | F1: 79.9 -> 82.3 | Improved |
| add knowledge distillation (ablation) | soft movement pruning · QQP dev, 3% weights | accuracy: 89.3 -> 89.1 | No clear effect |
| use global instead of local Top-v mask selection (ablation) | movement pruning · SQuAD, 3% weights | F1: +2.3 F1 | Mixed |
Closest papers by meaning
- Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision (ICML 2025 · similarity 0.90)
- The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws (ICLR 2025 · similarity 0.89)
- PAT: Pruning-Aware Tuning for Large Language Models (AAAI 2025 · similarity 0.89)
- SPP: Sparsity-Preserved Parameter-Efficient Fine-Tuning for Large Language Models (ICML 2024 · similarity 0.89)
- A Simple and Effective Pruning Approach for Large Language Models (ICLR 2024 · similarity 0.89)
- Fluctuation-Based Adaptive Structured Pruning for Large Language Models (AAAI 2024 · similarity 0.88)
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.
