Research map / Architectures and efficiency / Pruning and efficient tuning
A Fast Post-Training Pruning Framework for Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Prune heads and filters using Fisher-guided search, rearrangement, and mask tuning | BERT-base · SST-2 at 60% relative FLOPs | accuracy: 93.6 -> 92.5 | Hurt |
| Prune heads and filters using Fisher-guided search, rearrangement, and mask tuning | BERT-base · SST-2 on NVIDIA V100, batch size 256 | latency speedup: 1.56× | Helped |
| Add Fisher mask rearrangement (ablation) | BERT-base after mask search · GLUE and SQuAD at 60% FLOPs | accuracy or task score: 81.21 -> 81.81; 89.99 -> 90.08; 88.38 -> 88.77; 92.13 -> 92.09; 87.10 -> 87.68; 83.14 -> 83.23; 82.66 -> 84.47; 71.12 -> | Mixed |
| Add layer-wise mask tuning (ablation) | BERT-base after mask rearrangement · GLUE and SQuAD at 60% FLOPs | accuracy or task score: 81.81 -> 82.51; 90.08 -> 90.35; 88.77 -> 90.06; 92.09 -> 92.49; 87.68 -> 88.00; 83.23 -> 85.27; 84.47 -> 86.72; 72.38 -> | Helped |
| Use a larger calibration sample, from 2K to 4K or 8K examples (ablation) | Post-training pruning framework · Four GLUE datasets with more than 64K training examples | accuracy: up to 0.4% | Helped |
| Use Fisher importance instead of weight magnitude or gradient-based importance (ablation) | Mask search · GLUE and SQuAD pruning experiments | accuracy | Helped |
| Use mask search and rearrangement instead of uniform Fisher pruning (ablation) | BERT-base with mask tuning · GLUE and SQuAD | accuracy: up to 8.6% | Helped |
| Use magnitude-based or gradient-based pruning criteria instead of Fisher-based search (ablation) | BERT-base · Low-sparsity pruning regimes | accuracy | Hurt |
Closest papers by meaning
- Resource-Efficient Transformer Pruning for Finetuning of Large Models (CVPR 2024 · similarity 0.91)
- CORP: Closed-Form One-shot Representation-Preserving Structured Pruning for Transformers (NeurIPS 2026 · similarity 0.90)
- Adaptive Pruning of Pretrained Transformer via Differential Inclusions (ICLR 2025 · similarity 0.90)
- The Need for Speed: Pruning Transformers with One Recipe (ICLR 2024 · similarity 0.89)
- Zero-TPrune: Zero-Shot Token Pruning through Leveraging of the Attention Graph in Pre-Trained Transformers (CVPR 2024 · similarity 0.89)
- Numerical Pruning for Efficient Autoregressive Models (AAAI 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.