Research map / Architectures and efficiency / Pruning and efficient tuning
Pruning as a Cooperative Game: Surrogate-Assisted Layer Contribution Estimation for Large Language Models
Xuan Ding, Pengyu Tong, Ranjie Duan, Yunjian Zhang, Rui Sun, Yao Zhu
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 |
|---|---|---|---|
| prune after 4-bit quantization instead of before (3 layers) (ablation) | Shapley layer pruning + GPTQ · WikiText2, LLaMA-2-7B | perplexity: 15.4001 -> 14.9263 | Improved |
| prune after 4-bit quantization instead of before (6 layers) (ablation) | Shapley layer pruning + GPTQ · WikiText2, LLaMA-2-7B | perplexity: 19.7741 -> 18.5761 | Improved |
| split pruning around quantization (3 + 3 layers) vs pruning 6 layers first (ablation) | Shapley layer pruning + GPTQ · WikiText2, LLaMA-2-7B | perplexity: 19.7741 -> 19.4676 | Improved |
| re-estimate with joint pair deletion (Optimal Combo) instead of static two least important (ablation) | layer deletion · BookCorpus PPL, LLaMA-2-7B | perplexity: 15.4535 -> 15.4279 | Improved |
| use Shapley-surrogate pruning vs ShortGPT, remove 9 layers | Meta-LLaMA-3-8B · WikiText2 | perplexity: 2549.7485 -> 45.2635 | Improved |
| use Shapley-surrogate pruning vs SLEB, remove 12 layers | Meta-LLaMA-3-8B · WikiText2 | perplexity: 126.9445 -> 304.5240 | Hurt |
| use Shapley-surrogate pruning vs SLEB, remove 3 layers | LLaMA-2-7B · WikiText2 | perplexity: 14.2428 -> 14.6949 | Hurt |
| use Shapley-surrogate pruning vs SLEB, remove 12 layers | LLaMA-2-7B · WikiText2 | perplexity: 58.1194 -> 38.1157 | Improved |
| use Shapley-surrogate pruning vs Shortened-LLaMA, 6.1B | LLaMA-2-7B · zero-shot average | accuracy: 0.5816 -> 0.5782 | No clear effect |
| use Shapley-surrogate pruning vs SliceGPT, 5.5B | LLaMA-2-7B · zero-shot average | accuracy: 0.3865 -> 0.5227 | Improved |
| apply pruning to Mamba-2.8B instead of ShortGPT | Mamba-2.8B · WikiText2 | perplexity: 4074.4865 -> 31.1091 | Improved |
| test surrogate on masks differing from training masks (ks 16,14,12,21,18) (ablation) | surrogate network · LLaMA-2-7B mask prediction | R2: 0.9464 -> 0.2466 | Hurt |
Closest papers by meaning
- Discovering Sparsity Allocation for Layer-wise Pruning of Large Language Models (NeurIPS 2024 · similarity 0.91)
- Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation (NeurIPS 2025 · similarity 0.91)
- SlimLLM: Accurate Structured Pruning for Large Language Models (ICML 2025 · similarity 0.90)
- AlphaPruning: Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models (NeurIPS 2024 · similarity 0.90)
- SlimGPT: Layer-wise Structured Pruning for Large Language Models (NeurIPS 2024 · similarity 0.90)
- SkipGPT: Each Token is One of a Kind (ICML 2025 · similarity 0.90)
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.
