Research map / Architectures and efficiency / Pruning and efficient tuning
Sketch-Based Low-Rank Model Merging with Shared Circulant Transforms
Zhiming Zhang, Rong Yin, Xiaoshuai Hao, Hansong Zhang, Hao Peng, Yong Liu, Can Ma, Dan Meng
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 shared circulant transform before sketching (ablation) | CircuMerge · DARE-TIES, ViT-B/32 vision merge | avg normalized accuracy: 57.92 -> 69.77 | Improved |
| use circulant instead of dense random invertible transform (ablation) | CircuMerge · DARE-TIES, ViT-B/32 vision merge | avg normalized accuracy: 62.49 -> 69.77 | Improved |
| use circulant instead of dense random invertible transform (ablation) | CircuMerge · DARE-TIES, ViT-B/32 vision merge | merge time (s): 5.48 -> 3.72 | Improved |
| add random invertible shared transform vs none (ablation) | CircuMerge · DARE-TIES, ViT-B/32 vision merge | avg normalized accuracy: 57.92 -> 62.49 | Improved |
| increase sketch size s beyond 8 (ablation) | CircuMerge · NLI and vision suites | avg normalized accuracy | Hurt |
| merge in sketch space with CircuMerge instead of core space (DARE-TIES) | DARE-TIES · 6-task NLI (Llama 3 8B LoRA merge) | avg normalized accuracy: 93.32 -> 93.64 | Improved |
| merge in sketch space with CircuMerge instead of core space (TSV) | TSV · 6-task NLI (Llama 3 8B LoRA merge) | avg normalized accuracy: 94.16 -> 94.13 | No clear effect |
| merge in sketch space with CircuMerge instead of core space (Iso-C) | Iso-C · 6-task NLI (Llama 3 8B LoRA merge) | avg normalized accuracy: 90.33 -> 89.37 | Hurt |
| merge in sketch space with CircuMerge instead of core space (TSV+Iso-C) | TSV+Iso-C · 8-task vision (ViT-B/32 LoRA merge) | avg normalized accuracy: 76.43 -> 77.53 | Improved |
| merge in sketch space with CircuMerge instead of core space (TIES) | TIES · 8-task vision (ViT-B/32 LoRA merge) | avg normalized accuracy: 68.63 -> 68.39 | No clear effect |
| CircuMerge instead of KnOTS merge time | TIES · 6-task NLI (Llama 3 8B LoRA merge) | merge time (s): 2963 -> 3.7 | Improved |
Closest papers by meaning
- Accurate and Efficient Low-Rank Model Merging in Core Space (NeurIPS 2025 · similarity 0.90)
- Crowded in B-Space: Calibrating Shared Directions for LoRA Merging (NeurIPS 2026 · similarity 0.90)
- No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces (ICML 2025 · similarity 0.89)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.88)
- IterIS: Iterative Inference-Solving Alignment for LoRA Merging (CVPR 2025 · similarity 0.88)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.87)
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.
