Research map / Architectures and efficiency / Quantization
Prune-then-Quantize or Quantize-then-Prune? Understanding the Impact of Compression Order in Joint Model Compression
Minjun Kim, Jaehyeon Choi, Hyunwoo Yang, Jongjin Kim, Jinho Song, U Kang
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 |
|---|---|---|---|
| apply stronger compression after weaker (quantize first, then prune) as the CER gap grows | SparseGPT + QuaRot · WikiText-2 (LLaMA 2 7B/13B, LLaMA 3 8B) | compression order advantage | Improved |
| applying the progressive-intensity ordering to vision models (pruning + quantization) | ResNet-18 and DeiT-Base · ImageNet | accuracy | Improved |
| apply stronger pruning later in multi-stage compression | SparseGPT + QuaRot · WikiText-2 (LLaMA 3 8B) | perplexity | Improved |
| progressive (8 to 2 bit) instead of regressive bit allocation in mixed-precision quantization | HAWQ-V2 · ImageNet (ResNet-18) | accuracy | Improved |
| order of pruning and parameter sharing | Basis Sharing + magnitude pruning · LLaMA 2 7B | perplexity | Improved |
| add Hadamard rotation before SparseGPT pruning (ablation) | SparseGPT · WikiText-2 (LLaMA 3 8B, ratio 0.4) | perplexity: 8.477 -> 98.213 | Hurt |
| add Hadamard rotation before SLEB pruning (ablation) | SLEB · WikiText-2 (LLaMA 3 8B, ratio 0.4) | perplexity: 92.848 -> 93.260 | No clear effect |
| unstructured (SparseGPT) vs structured (SLEB) pruning in order interference at 4-bit | SparseGPT, SLEB · WikiText-2 (LLaMA 3 8B) | order advantage: -49.899 vs -9.379 | Mixed |
| add post-quantization LoRA | SparseGPT + RTN · LLaMA 3 8B | perplexity | Improved |
Closest papers by meaning
- In defense of parameter sharing for model-compression (ICLR 2024 · similarity 0.88)
- Accuracy is Not All You Need (NeurIPS 2024 · similarity 0.88)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.88)
- Unified Scaling Laws for Compressed Representations (NeurIPS 2025 · similarity 0.88)
- Breaking the Compression Ceiling: Data-Free Pipeline for Ultra-Efficient Delta Compression (NeurIPS 2025 · similarity 0.88)
- SLiM: One-shot Quantization and Sparsity with Low-rank Approximation for LLM Weight Compression (ICML 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.
