Research map / Architectures and efficiency / Pruning and efficient tuning
Sliding-Window Merging for Compacting Patch-Redundant Layers in LLMs
Xuan Ding, Rui Sun, Yunjian Zhang, Xiu Yan, Yueqi Zhou, Kaihao Huang, Suzhong Fu, Angelica I Aviles-Rivero, Chuanlong Xie, 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 |
|---|---|---|---|
| merge layers by parameter differences (SWM) instead of deleting them or averaging (ablation) | LLaMA2-7B depth pruning · LLaMA2-7B, threshold 0.75 | layers retained at equal threshold: 23 layers vs 26 (Delete) vs 29 (Average) | Improved |
| SWM vs best baseline at 20% pruning | SWM · LLaMA2-7B zero-shot avg | average accuracy: 59.268 -> 61.944 (+2.676) | Improved |
| SWM vs best baseline at 35% pruning | SWM · Vicuna-7B zero-shot avg | average accuracy: 54.596 -> 56.250 | Improved |
| SWM vs LLM-Pruner at 35% pruning | SWM · LLaMA2-7B zero-shot avg | average accuracy: 52.366 -> 51.869 | Hurt |
| add LoRA retraining after pruning (ablation) | SWM · LLaMA2-7B 20% | average accuracy: 57.360 -> 61.944 | Improved |
| combine SWM depth stage (18-91%) with LLM-Pruner width stage at 35% total sparsity (ablation) | SWM + LLM-Pruner · LLaMA2-7B | average accuracy | Improved |
| vary calibration dataset and sample size (ablation) | SWM · LLaMA2-7B HellaSwag | accuracy: 62.179 to 63.730 | No clear effect |
| use depth pruning instead of width pruning | SWM/SLEB/Shortened-LLM · LLaMA2-7B 20% | latency: 5.655 -> 2.339 (LLM-Pruner vs SWM) | Improved |
| lower similarity threshold (ablation) | SWM · LLaMA2-7B | compression vs performance | Mixed |
Closest papers by meaning
- Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization (NeurIPS 2025 · similarity 0.87)
- Collaborative Multi-Mode Pruning for Vision-Language Models (CVPR 2026 · similarity 0.87)
- MUSTAFAR: Promoting Unstructured Sparsity for KV Cache Pruning in LLM Inference (NeurIPS 2025 · similarity 0.87)
- SLEB: Streamlining LLMs through Redundancy Verification and Elimination of Transformer Blocks (ICML 2024 · similarity 0.86)
- Layer-Adaptive State Pruning for Deep State Space Models (NeurIPS 2024 · similarity 0.86)
- Let LLM Tell What to Prune and How Much to Prune (ICML 2025 · similarity 0.86)
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.
