Research map / Large language models / Instruction and fine-tuning
Reassessing Layer Pruning in LLMs: New Insights and Methods
Yao Lu, Hao Cheng, Yujie Fang, Zeyu Wang, Jiaheng Wei, Dongwei Xu, Qi Xuan, Zhaowei 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 |
|---|---|---|---|
| use reverse-order layer selection instead of PPL (ablation) | LoRA-tuned pruned LLMs · 25% pruning, 4 models | Avg accuracy: outperforms the second-best PPL metric by 5.30% | Improved |
| use reverse-order instead of Taylor on Llama-3.1-8B-It (ablation) | pruned Llama-3.1-8B-It · 25% pruning, LoRA | Avg accuracy: 0.4796 -> 0.5268 | Improved |
| use reverse-order instead of Taylor on Gemma2-2B-It (ablation) | pruned Gemma2-2B-It · 25% pruning, LoRA | Avg accuracy: 0.5073 -> 0.5032 | No clear effect |
| use partial-layer fine-tuning (lm_head+last three layers) instead of LoRA (ablation) | pruned Llama-3.1-8B-It · 25% reverse-order pruning | Avg accuracy: 0.5268 -> 0.5807 | Improved |
| use partial-layer fine-tuning (lm_head+last three layers) instead of LoRA (ablation) | pruned Qwen1.5-7B · 25% reverse-order pruning | Avg accuracy: 0.5037 -> 0.5445 | Improved |
| use partial-layer fine-tuning (lm_head+last two layers) instead of LoRA (ablation) | pruned Vicuna-7B-v1.5 · 25% reverse-order pruning | Avg accuracy: 0.4909 -> 0.5060 | Improved |
| use QLoRA instead of LoRA (ablation) | pruned Llama-3.1-8B-It · 25% reverse-order pruning | Avg accuracy: 0.5268 -> 0.5171 | Hurt |
| fine-tune more last layers (lm_head only to lm_head+last three layers) (ablation) | pruned Llama-3.1-8B-It · 25% reverse-order pruning | Avg accuracy: 0.5599 -> 0.5807 | Improved |
| use partial-layer fine-tuning instead of LoRA on the unpruned model (ablation) | Llama-3.1-8B-It dense · Full-model fine-tuning | Avg accuracy: 0.6354 -> 0.6337 | No clear effect |
| use iterative pruning (1:4:8) instead of one-shot with LoRA (ablation) | pruned Llama-3.1-8B-It, reverse-order · 25% pruning | Avg accuracy: 0.5268 -> 0.5455 | Improved |
| use iterative pruning (1:1:8) instead of one-shot with Taylor and partial-layer fine-tuning (ablation) | pruned Llama-3.1-8B-It, Taylor · 25% pruning | Avg accuracy: 0.5701 -> 0.3753 | Hurt |
| use iterative pruning (1:1:6) instead of one-shot with Taylor and LoRA (ablation) | pruned Gemma2-2B-It · 25% pruning | Avg accuracy: 0.5073 -> 0.4947 | Hurt |
Closest papers by meaning
- Streamlining Redundant Layers to Compress Large Language Models (ICLR 2025 · similarity 0.93)
- SlimLLM: Accurate Structured Pruning for Large Language Models (ICML 2025 · similarity 0.93)
- Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation (NeurIPS 2025 · similarity 0.93)
- SlimGPT: Layer-wise Structured Pruning for Large Language Models (NeurIPS 2024 · similarity 0.92)
- DLP: Dynamic Layerwise Pruning in Large Language Models (ICML 2025 · similarity 0.92)
- AlphaPruning: Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models (NeurIPS 2024 · similarity 0.92)
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.
