Research map / Architectures and efficiency / Pruning and efficient tuning
Deterministic Differentiable Structured Pruning for Large Language Models
Weiyu Huang, Pengle Zhang, Xiaolu Zhang, JUN ZHOU, Jun Zhu, Jianfei Chen
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 |
|---|---|---|---|
| remove sampling noise from hard-concrete masks and add binarization loss (deterministic HC) (ablation) | hard-concrete mask learning · LLaMA-7B 20% sparsity | mean zero-shot accuracy: 59.95 -> 61.74 | Improved |
| add expanded ReLU mask parameterization (decouple forward gate from sparsity regularizer) (ablation) | Det. HC · LLaMA-7B 20% sparsity | mean zero-shot accuracy: 61.74 -> 63.92 | Improved |
| add knowledge distillation from dense model (ablation) | Det. HC + EM · LLaMA-7B 20% sparsity | mean zero-shot accuracy: 63.92 -> 64.13 | Improved |
| add distillation on DeepSeekMoE-16B (ablation) | Det. HC + EM · DeepSeekMoE-16B 20% sparsity | mean zero-shot accuracy: 61.80 -> 61.84 | No clear effect |
| use layer-wise instead of global sparsity (ablation) | DDP · LLaMA-7B 20% sparsity | mean accuracy change: -0.48 pp | Hurt |
| use expert-wise instead of global sparsity (ablation) | DDP · DeepSeekMoE-16B 20% sparsity | mean accuracy change: -1.97 pp | Hurt |
| train masks on C4 instead of FineWeb-Edu (ablation) | DDP · LLaMA-7B 20% sparsity | mean accuracy change: -1.14 pp | Hurt |
| train masks on LaMini instruction data instead of FineWeb-Edu (ablation) | DDP · DeepSeekMoE-16B 20% sparsity | mean accuracy change: -0.87 pp | Hurt |
| increase mask-learning tokens beyond 60M (ablation) | DDP · LLaMA-7B and DeepSeekMoE-16B 20% sparsity | perplexity and accuracy | No clear effect |
| add LoRA recovery tokens to training-free pruning instead of mask optimization (ablation) | LoRAP on LLaMA-2-7B · LLaMA-2-7B 20% sparsity | mean zero-shot accuracy: 59.44 -> 62.98 at 120M tokens vs DDP 64.82 at 30M | Mixed |
| use DDP instead of one-shot pruning with fine-tuning | pruned LLMs · DeepSeekMoE-16B 60% sparsity | mean accuracy: 51.62 -> 58.18 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SlimLLM — SlimLLM: Accurate Structured Pruning for Large Language Models (ICML 2025) · better on 4 of 4 shared results
- LoRAP — LoRAP: Transformer Sub-Layers Deserve Differentiated Structured Compression for Large Language Models (ICML 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization (NeurIPS 2025 · similarity 0.92)
- Fluctuation-Based Adaptive Structured Pruning for Large Language Models (AAAI 2024 · similarity 0.91)
- Discovering Sparsity Allocation for Layer-wise Pruning of Large Language Models (NeurIPS 2024 · similarity 0.90)
- An Efficient Pruner for Large Language Model with Theoretical Guarantee (ICML 2025 · similarity 0.90)
- MaskLLM: Learnable Semi-Structured Sparsity for Large Language Models (NeurIPS 2024 · similarity 0.90)
- SparseLLM: Towards Global Pruning of Pre-trained Language Models (NeurIPS 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.
