Research map / Architectures and efficiency / Pruning and efficient tuning
Label-Free Cross-Task LoRA Merging with Null-Space Compression
Wonyoung Lee, Wooseong Jeong, Kuk-Jin Yoon
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 |
|---|---|---|---|
| use NSC null-space objective instead of entropy surrogate (AdaMerging) for NLI merging | LoRA merging · 6 NLI tasks (LLaMA-3 8B) | avg normalized accuracy (%): 89.6 -> 92.3 | Improved |
| use NSC instead of gradient-free merging (TA/TIES/DARE-TIES) on dense vision tasks | ViT-B LoRA merging · 20 vision tasks | avg normalized performance (%): 77.2 -> 92.0 (TA vs NSC) | Improved |
| NSC instead of AdaMerging Single Token on VLM merging | LLaVA-1.5-7B LoRA merging · 6 VLM tasks | avg normalized score (%): 80.9 -> 82.7 | Improved |
| NSC instead of AdaMerging Full Token on VLM merging | LLaVA-1.5-7B LoRA merging · 6 VLM tasks | avg normalized score (%): 82.4 -> 82.7 | No clear effect |
| NSC instead of AdaMerging Full Token for optimization cost | LLaVA-1.5-7B LoRA merging · LLaVA 6 tasks | total time (min): 104.0 -> 13.3 | Improved |
| NSC instead of gradient-free merging on unseen tasks | ViT-B LoRA merging · 20 vision tasks, unseen split | avg normalized performance (%): 76.8 -> 87.1 (TA vs NSC) | Improved |
| restrict NSC objective to only 1 transformer block instead of 3-6 (ablation) | NSC objective (O projection) · 8 image classification tasks (CLIP ViT-B/32) | accuracy (%): 84.6 -> 82.6 | Hurt |
| activate all 12 blocks instead of 3-6 for the NSC objective (ablation) | NSC objective (O projection) · 8 image classification tasks (CLIP ViT-B/32) | accuracy (%): 84.9 (6 blocks) -> 84.5 (12 blocks) | No clear effect |
| target only O projection instead of QKVO (ablation) | NSC objective · 8 image classification tasks (CLIP ViT-B/32) | accuracy (%): 83.7 -> 84.5 (12 blocks) | No clear effect |
| gradient-free low-rank SVD or Linear merging instead of NSC | ViT-B LoRA merging · 20 vision tasks | avg normalized performance (%): SVD 83.0, Linear 74.3 vs NSC 92.0 | Mixed |
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.
- RobustMerge — RobustMerge: Parameter-Efficient Model Merging for MLLMs with Direction Robustness (NeurIPS 2025) · better on 5 of 5 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
- Accurate and Efficient Low-Rank Model Merging in Core Space (NeurIPS 2025 · similarity 0.90)
- IterIS: Iterative Inference-Solving Alignment for LoRA Merging (CVPR 2025 · similarity 0.88)
- NOLA: Compressing LoRA using Linear Combination of Random Basis (ICLR 2024 · similarity 0.88)
- Crowded in B-Space: Calibrating Shared Directions for LoRA Merging (NeurIPS 2026 · similarity 0.87)
- Merging LoRAs like Playing LEGO: Pushing the Modularity of LoRA to Extremes Through Rank-Wise Clustering (ICLR 2025 · similarity 0.87)
- Localizing Task Information for Improved Model Merging and Compression (ICML 2024 · 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.
