Research map / Architectures and efficiency / Pruning and efficient tuning
SALR: Sparsity-Aware Low-Rank Representation for Efficient Fine-Tuning of Large Language Models
Longteng Zhang, Sen Wu, Shuai Hou, Zhengyu Qing, Zhuo Zheng, Danning Ke, Qihong Lin, Qiang Wang, Shaohuai Shi, Xiaowen Chu
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 |
|---|---|---|---|
| make the SVD residual trainable instead of frozen (ablation) | SALR · MMLU | accuracy: 66.8 -> 68.2 (Llama3-8B); 54.2 -> 56 (Llama2-7B) | Improved |
| prune only the frozen base weights with a static mask rather than pruning LoRA modules (LoSA style) | LoRA pruning · GSM8K, 50% sparsity | accuracy: 71.4 -> 79.5 (Llama3-8B) | Improved |
| add low-rank adapter capturing pruned residual via truncated SVD | static-pruned LoRA · GSM8K / MMLU, 50% sparsity | accuracy: SALR 56.7 vs LoSA 34.2 GSM8K (Llama2-7B) | Improved |
| increase sparsity from 10% to 50% (ablation) | SALR · GSM8K | accuracy: 79.5 -> 80.1 (30%) -> 79.5 (50%) | No clear effect |
| add NF4 quantization on 20% sparse SALR (ablation) | SALR · GSM8K | accuracy: 71 -> 70.4 (DeepSeek-V2-Lite); 79.2 -> 79.2 (Mixtral-8x7B) | No clear effect |
| prune only base weights and use fused low-rank adapters instead of LoSA dense update | SALR fine-tuning · Llama3-8B fine-tuning, 50% sparsity | GPU memory (GB): 27.1 -> 19.2 | Improved |
| use SALR instead of LoSA under 2:4 sparsity | Llama3-8B inference · GSM8K, RTX4090 | accuracy: accuracy 69.4 -> 78.9; speedup 1.9x -> 1.7x | Mixed |
| use SALR instead of LoRA on Llama3-8B | SALR · MMLU, 50% sparsity | accuracy: 69.2 -> 68.2 | Hurt |
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.
- LoSA — LoSA: Locality Aware Sparse Attention in Diffusion Language Models (ICML 2026) · better on 8 of 9 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 Low-Rank Sparse Adaptation for Large Language Models (ICLR 2025 · similarity 0.92)
- PAT: Pruning-Aware Tuning for Large Language Models (AAAI 2025 · similarity 0.91)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.91)
- SPP: Sparsity-Preserved Parameter-Efficient Fine-Tuning for Large Language Models (ICML 2024 · similarity 0.91)
- APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference (ICML 2024 · similarity 0.91)
- The Sparsity Whisperer (NeurIPS 2026 · 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.
