Research map / Architectures and efficiency / RNNs and normalization
Towards Minimizing Feature Drift in Model Merging: Layer-wise Task Vector Fusion for Adaptive Knowledge Integration
Read the paperCheck your idea against 108,967 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 LOT Merging, a training-free model merging method that minimizes layer-wise feature drift via closed-form per-layer task-vector solutions using a few exemplars per task.
Problem
Multi-task model merging loses performance versus fine-tuned experts; parameter-level methods lag the upper bound and task-loss methods need costly secondary training. The paper targets performance degradation as caused by feature drift introduced by merging.
Approach
Merge task vectors layer by layer by minimizing squared feature drift on small exemplar sets; this is a convex quadratic problem with closed-form solutions for linear weights (pseudo-inverse of summed Gram matrices), normalization scales (weighted average) and biases (mean of task vectors).
Key results
- On ViT-B/32 eight vision tasks, LOT Merging reaches 82.7% average accuracy vs 78.3% for the next-best training-free method CAT Merging (about 4.4 points gain).
- On ViT-L/14 eight vision tasks, LOT Merging reaches 90.5% average accuracy, 0.9 points above CAT Merging (89.6%).
- On BLIP vision-language tasks, LOT Merging is best on 5 of 6 tasks, e.g., ScienceQA 48.24 vs 46.36 for CAT Merging.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LOT Merging instead of CAT Merging | ViT-B/32 model merging · Eight vision tasks (Avg Acc) | accuracy: 78.3 -> 82.7 | Helped |
| use LOT Merging instead of CAT Merging | ViT-L/14 model merging · Eight vision tasks (Avg Acc) | accuracy: 89.6 -> 90.5 | Helped |
| use LOT Merging instead of CAT Merging | BLIP model merging · OKVQA | accuracy: 44.07 -> 38.35 | Hurt |
| add merged linear weights (ablation) | LOT Merging on ViT-B/32 · Eight vision tasks (Avg Acc) | accuracy: 49.8 -> 82.7 | Helped |
| add merged scalers (ablation) | LOT Merging on ViT-B/32 · Eight vision tasks (Avg Acc) | accuracy: 81.9 -> 82.7 | Helped |
| add merged bias (ablation) | LOT Merging on ViT-B/32 · Eight vision tasks (Avg Acc) | accuracy: 82.4 -> 82.7 | Helped |
| add merged linear weights (ablation) | LOT Merging on ViT-L/14 · Eight vision tasks (Avg Acc) | accuracy: 65.6 -> 90.5 | Helped |
| add merged scalers (ablation) | LOT Merging on ViT-L/14 · Eight vision tasks (Avg Acc) | accuracy: 90.3 -> 90.5 | Helped |
| add merged bias (ablation) | LOT Merging on ViT-L/14 · Eight vision tasks (Avg Acc) | accuracy: 90.2 -> 90.5 | Helped |
Surprising findings
- Direct parameter merging (solving for merged weights rather than task vectors) degrades strongly with few exemplars, mostly due to altering pre-trained weights.
Limitations
- Requires a small exemplar set, so not strictly data-free.
- Extending to layers with more complex operations (e.g., exponential functions) may need adaptation.
- Lags compression-based methods such as WEMoE and EMR-Merging in average accuracy on ViT-B/32 and ViT-L/14.
Experimental setup
Datasets: SUN397 Stanford Cars RESISC45 EuroSAT SVHN GTSRB MNIST DTD COCO Caption Flickr30k Caption TextCaps OKVQA
Benchmarks: SUN397 Stanford Cars RESISC45 EuroSAT SVHN GTSRB MNIST DTD COCO Caption Flickr30k Caption TextCaps OKVQA
Metrics: accuracy CIDEr
Baselines: Weight Averaging Fisher Merging RegMean Task Arithmetic Ties-Merging TATR Consensus Merging AWD Merging PCB Merging CAT Merging WEMoE EMR-Merging
Builds on: Task Arithmetic RegMean Sun et al. layer-wise negative transfer bound
Closest papers by meaning
- Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · similarity 0.91)
- CAT Merging: A Training-Free Approach for Resolving Conflicts in Model Merging (ICML 2025 · similarity 0.91)
- Modeling Multi-Task Model Merging as Adaptive Projective Gradient Descent (ICML 2025 · similarity 0.90)
- Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors (ICML 2025 · similarity 0.90)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.89)
- Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging (NeurIPS 2025 · similarity 0.88)
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.