Research map / Architectures and efficiency / Pruning and efficient tuning
SliceFine: The Universal Winning-Slice Hypothesis for Pretrained Networks
Md Kowsher, Ali O. Polat, Ehsan Mohammady Ardehaly, Mehrdad Salehi, Zia Ghiasi, Prasanth Murali, Chen 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 |
|---|---|---|---|
| increase slice rank from 1 to 5 (alternating row-column) (ablation) | SliceFine · Math reasoning avg (LLaMA-3) | average accuracy: 81.55 -> 82.13 | Improved |
| increase slice rank from 1 to 5 (row slices) (ablation) | SliceFine · VTAB-1K image avg | average accuracy: 88.15 -> 88.85 | Improved |
| increase slice rank beyond a moderate range (ablation) | SliceFine · downstream accuracy | validation accuracy | Hurt |
| choose slice position at different places in the weight matrix (ablation) | SliceFine (rank 5) · Figure 3(b) | accuracy: within ±1% of anchor | No clear effect |
| select slices by Wanda importance (most/least important, mixed, random) (ablation) | SliceFine · Figure 3(c) | accuracy | No clear effect |
| use slices from 'bad' LTH subnetworks instead of 'good' ones (ablation) | SliceFine · Figure 3(d) | accuracy | No clear effect |
| replace structured row/column slices with unstructured random masks (ablation) | SliceFine · Tables 9 and 10 | accuracy | No clear effect |
| use unstructured random masks instead of structured slices (ablation) | SliceFine · training efficiency | memory and training time | Hurt |
| switch the active slice too frequently (ablation) | SliceFine · switching interval ablation | accuracy | Hurt |
| use dynamic slice position instead of static (ablation) | SliceFine · slice position ablation | accuracy | Improved |
| prune a larger fraction of frozen backbone weights (ablation) | SliceFine · backbone pruning (Figure 9) | accuracy | Hurt |
| use SliceFine instead of LoRA/HRA/AdaLoRA for training | PEFT on ViT/VideoMAE/RoBERTa · throughput | iterations/s: 2.05 vs 1.78 (LoRA) and 1.62 (HRA) | 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.
- MiSS — MiSS: Revisiting the Trade-off in LoRA with an Efficient Shard-Sharing Structure (ICLR 2026) · better on 2 of 2 shared results
- VeRA — VeRA: Vector-based Random Matrix Adaptation (ICLR 2024) · 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
- Linearization Explains Fine-Tuning in Large Language Models (NeurIPS 2025 · similarity 0.85)
- UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory (CVPR 2024 · similarity 0.85)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.85)
- Synergy and Diversity in CLIP: Enhancing Performance Through Adaptive Backbone Ensembling (ICLR 2025 · similarity 0.85)
- Pay Attention to Small Weights (NeurIPS 2025 · similarity 0.85)
- SharpZO: Hybrid Sharpness-Aware Vision Language Model Prompt Tuning via Forward-Only Passes (NeurIPS 2025 · 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.
