Research map / Architectures and efficiency / Pruning and efficient tuning
Breaking Information Islands in Sparse Tuning via Small-World Connectivity
Xianchao Guan, Zijun Xiong, Yifeng Wang, Yubo Cui, Yaowei Wang, Guosen Xie, Xin Li, Zheng Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a sparse-tuning approach for LLM adaptation that uses small-world connectivity to avoid information islands among feature dimensions.
Problem
Sparse tuning of large language models depends on the feature-interaction topology induced by the sparse update; limited or uneven cross-dimensional interactions can isolate feature dimensions into information islands that hinder task-specific information mixing.
Approach
Structures the sparse update pattern with small-world connectivity so that feature dimensions are well interconnected, breaking information islands while keeping the trainable parameter count sparse.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add fixed scattering permutation to block-diagonal update (SWIFT vs DiaBlo) | block-diagonal sparse tuning · Commonsense reasoning, LLaMA3-8B | average accuracy (%): 87.00 -> 88.05 | Improved |
| add fixed scattering permutation to block-diagonal update (SWIFT vs DiaBlo) | block-diagonal sparse tuning · NLG, LLaMA2-7B | average score (%): 31.53 -> 32.56 | Improved |
| add fixed scattering permutation to block-diagonal update (SWIFT vs DiaBlo) | block-diagonal sparse tuning · Image classification, CLIP ViT-B/32 | average accuracy (%): 85.07 -> 86.78 | Improved |
| use uniform scattering permutation instead of random permutation (ablation) | SWIFT · Image classification, CLIP ViT-B/32 | average accuracy (%): +0.77 percentage points | Improved |
| use uniform scattering permutation instead of diagonal cyclic permutation (ablation) | SWIFT · Image classification, CLIP ViT-B/32 | average accuracy (%): +1.71 percentage points | Improved |
| increase update density beyond N=64 (N=32, 3.13%) (ablation) | SWIFT · Commonsense reasoning, LLaMA3-8B | average accuracy (%) | No clear effect |
| use N=64 blocks (density 1.56%) as default (ablation) | SWIFT · Commonsense reasoning, LLaMA3-8B | average accuracy (%): 88.05 | Improved |
| apply SWIFT on attention projections only vs DiaBlo (ablation) | SWIFT · Commonsense reasoning, LLaMA3-8B | average accuracy (%): +1.24 percentage points | Improved |
| apply SWIFT on feed-forward projections only vs DiaBlo (ablation) | SWIFT · Commonsense reasoning, LLaMA3-8B | average accuracy (%): +0.93 percentage points | Improved |
| use block-diagonal structure with scattering permutation for MRGC | SWIFT support · MRGC, d=4096, K=64 | MRGC: 0.00 -> 1.00 | Improved |
| use block-diagonal permuted support instead of random masks | SWIFT support · MRGC, d=4096, K=64 | MRGC: 0.41 -> 1.00 | Improved |
| use SWIFT instead of LoRA | LLaMA3-8B · Commonsense reasoning | average accuracy (%): 85.01 -> 88.05 | 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.
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 2 of 2 shared results
- S2FT — S2FT: Parameter-Efficient Fine-Tuning in Sparse Spectrum Domain (CVPR 2026) · better on 1 of 1 shared result
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 0.86)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.85)
- SWIFT: A Scalable Lightweight Infrastructure for Fine-Tuning (AAAI 2025 · similarity 0.85)
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.85)
- CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models (ICML 2025 · similarity 0.85)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.85)
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.
