Research map / Architectures and efficiency / Pruning and efficient tuning
Alignment-Enhanced Integration of Connectivity and Spectral Sparse in Dynamic Sparse Training of LLM
Wenjing Wu, Yingtao Zhang, Jialin Zhao, Carlo Vittorio Cannistraci
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 alignment loss on top of activation adjustment (ablation) | CHTs + low-rank · OpenWebText LLaMA-60M, sparsity 0.9 | validation perplexity: 32.21 -> 31.77 | Improved |
| add activation adjustment to naive sum of CHTs and low-rank (ablation) | CHTs + low-rank · C4 LLaMA-60M, sparsity 0.9 | validation perplexity: 189.55 -> 39.66 | Improved |
| add activation adjustment to naive sum (ablation) | CHTs + low-rank · OpenWebText LLaMA-130M, sparsity 0.9 | validation perplexity: 119.35 -> 24.45 | Improved |
| add activation adjustment to naive sum (ablation) | CHTs + low-rank · OpenWebText LLaMA-60M, sparsity 0.7 | validation perplexity: 27.89 -> 29.94 | Hurt |
| add activation adjustment to naive sum (ablation) | CHTs + low-rank · OpenWebText LLaMA-60M, sparsity 0.8 | validation perplexity: 33.35 -> 29.42 | Improved |
| add full alignment scheme (activation + alignment loss) versus naive sum (ablation) | CHTs + low-rank · all models/datasets/sparsities (Wilcoxon) | validation perplexity: p-value 0.00049 | Improved |
| add alignment loss on top of activation adjustment (ablation) | CHTs + low-rank · all models/datasets/sparsities (Wilcoxon) | validation perplexity: p-value 0.00049 | Improved |
| add activation adjustment alone versus naive sum (ablation) | CHTs + low-rank · all models/datasets/sparsities (Wilcoxon) | validation perplexity: p-value 0.0093 | Improved |
| replace static sparse branch with dynamic CHTs branch plus alignment | SLTrain -> CHTsL · OpenWebText LLaMA-60M, sparsity 0.9 | validation perplexity: 33.90 -> 31.77 | Improved |
| add low-rank branch with alignment to CHTs | CHTs -> CHTsL · OpenWebText LLaMA-60M, sparsity 0.9 | validation perplexity: 33.03 -> 31.77 | Improved |
| add activation adjustment to static sparse plus low-rank (ablation) | static sparse + low-rank · OpenWebText LLaMA-130M, sparsity 0.7 | validation perplexity: 21.25 -> 41.51 | Hurt |
| add alignment loss on top of activation adjustment (ablation) | static sparse + low-rank · C4 LLaMA-130M, sparsity 0.7 | validation perplexity: 26.52 -> 26.36 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SLTrain — SLTrain: a sparse plus low rank approach for parameter and memory efficient pretraining (NeurIPS 2024) · better on 12 of 12 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
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.92)
- Dynamic Low-Rank Sparse Adaptation for Large Language Models (ICLR 2025 · similarity 0.90)
- Pruning Large Language Models with Semi-Structural Adaptive Sparse Training (AAAI 2025 · similarity 0.90)
- R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference (ICLR 2025 · similarity 0.89)
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 2025 · similarity 0.89)
- SparseLLM: Towards Global Pruning of Pre-trained Language Models (NeurIPS 2024 · similarity 0.89)
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.
