Research map / Architectures and efficiency / Pruning and efficient tuning
GradPruner: Gradient-guided Layer Pruning Enabling Efficient Fine-Tuning and Inference for LLMs
Wei Huang, Anda Cheng, Yinggui Wang
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 |
|---|---|---|---|
| add sign-aware layer merging after pruning 10 layers (3 merged) (ablation) | GradPruner, Llama3.1-8B FFT · 8-dataset avg | avg score: 0.741 -> 0.782 | Improved |
| add layer merging (2 merged) (ablation) | GradPruner, Llama3.1-8B FFT · 8-dataset avg | avg score: 0.767 -> 0.786 | Improved |
| add layer merging (3 merged) (ablation) | GradPruner, Llama3.1-8B LoRA · 8-dataset avg | avg score: 0.755 -> 0.786 | Improved |
| add layer merging (2 merged) (ablation) | GradPruner, Mistral-7B FFT · 8-dataset avg | avg score: 0.740 -> 0.770 | Improved |
| prune more than ~10 layers with layer pruning only (ablation) | GradPruner, Llama3.1-8B · downstream accuracy | accuracy | Hurt |
| merge only same-sign parameters instead of weight averaging (ablation) | GradPruner · downstream accuracy | accuracy | Improved |
| prune hidden-size kernels instead of layers (ablation) | GradPruner · downstream accuracy | accuracy | Hurt |
| accumulate gradients over 0.02% of steps instead of 1% (ablation) | GradPruner, Llama3.1-8B FFT · HellaSwag | accuracy: 0.939 -> 0.911 | Hurt |
| use 1% of WinoGrande training data instead of 100% (ablation) | GradPruner, Llama3.1-8B FFT · WinoGrande | accuracy: 0.861 -> 0.842 | Hurt |
| use GradPruner instead of Laco or SAT at 20% sparsity | GradPruner, Llama3.1-8B FFT · HellaSwag | accuracy: 0.942 vs 0.936 (Laco), 0.940 (SAT) | No clear effect |
| use GradPruner instead of other pruners at 40% sparsity | Llama3.1-8B FFT · 8-dataset avg | avg score: 0.734-0.750 -> 0.782 | Improved |
| prune 40% of parameters with GradPruner | Llama3.1-8B FFT · 8-dataset avg | avg score: 0.784 -> 0.782 | Hurt |
Closest papers by meaning
- APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference (ICML 2024 · similarity 0.91)
- Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization (NeurIPS 2025 · similarity 0.91)
- PAT: Pruning-Aware Tuning for Large Language Models (AAAI 2025 · similarity 0.91)
- Gradient Weight-normalized Low-rank Projection for Efficient LLM Training (AAAI 2025 · similarity 0.91)
- Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation (NeurIPS 2025 · similarity 0.91)
- SPP: Sparsity-Preserved Parameter-Efficient Fine-Tuning for Large Language Models (ICML 2024 · 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.
