Research map / Architectures and efficiency / Pruning and efficient tuning
A Win-win Deal: Towards Sparse and Robust Pre-trained Language Models
Yuanxin Liu, Fandong Meng, Zheng Lin, Jiangnan Li, Peng Fu, Yanan Cao, Weiping Wang, Jie Zhou
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 hard mask initialization with soft (ablation) | mask training at 90% sparsity · QQP | accuracy: 71.64 -> 77.08 | Improved |
| replace hard mask initialization with soft (ablation) | mask training at 90% sparsity · FEVER | accuracy: 49.56 -> 72.80 | Improved |
| replace hard mask initialization with soft (ablation) | mask training at 90% sparsity · MNLI | accuracy: 72.09 -> 72.63 | No clear effect |
| gradual sparsity increase 0.7~0.9 instead of fixed 90% (soft) (ablation) | mask training · MNLI | accuracy: 72.63 -> 76.84 | Improved |
| gradual sparsity increase 0.7~0.9 instead of fixed 90% (soft) (ablation) | mask training · HANS | accuracy: 52.82 -> 56.72 | Improved |
| gradual sparsity increase 0.7~0.9 instead of fixed 90% (soft) (ablation) | mask training · QQP | accuracy: 77.08 -> 79.49 | Improved |
| gradual sparsity increase 0.7~0.9 instead of fixed 90% (soft) (ablation) | mask training · FEVER | accuracy: 72.80 -> 79.01 | Improved |
| gradual sparsity increase 0.7~0.9 instead of fixed 90% (soft) (ablation) | mask training · PAWS-qqp | accuracy: 46.48 -> 46.59 | No clear effect |
| start gradual sparsity at 0.7 instead of 0.2 (ablation) | mask training at 90% sparsity · MNLI | accuracy: 73.61 -> 76.84 | Improved |
| gradual sparsity 50% to 70% instead of fixed 70% | mask training PoE · RoBERTa-base MNLI | accuracy: 83.48 -> 84.41 | Improved |
| gradual sparsity 50% to 70% instead of fixed 70% | mask training PoE · BERT-large MNLI | accuracy: 77.54 -> 84.83 | Improved |
| gradual sparsity 50% to 70% instead of fixed 70% | mask training PoE · BERT-large HANS | accuracy: 60.19 -> 70.18 | Improved |
Closest papers by meaning
- SparseLLM: Towards Global Pruning of Pre-trained Language Models (NeurIPS 2024 · similarity 0.89)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.88)
- Outlier Weighed Layerwise Sparsity (OWL): A Missing Secret Sauce for Pruning LLMs to High Sparsity (ICML 2024 · similarity 0.88)
- SPP: Sparsity-Preserved Parameter-Efficient Fine-Tuning for Large Language Models (ICML 2024 · similarity 0.88)
- MaskLLM: Learnable Semi-Structured Sparsity for Large Language Models (NeurIPS 2024 · similarity 0.88)
- Pruning Large Language Models with Semi-Structural Adaptive Sparse Training (AAAI 2025 · similarity 0.87)
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.
