Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Saliency-Aware Model Merging

About

Model merging aims to consolidate multiple task-specific models fine-tuned on different datasets into a unified architecture that performs cross-domain proficiency. Current data-free model merging methods often struggle to scale as they rely on simple parameter-level heuristics that ignore inter-layer dependencies and non-uniform distribution of expertise. This work proposes SA-Merging, which is built upon connectivity-based saliency formulations from structural pruning (e.g., SynFlow) and extends them to the data-free model merging setting. We define a saliency score over task vectors relative to a shared base model, and further introduce merge-aware modulation that incorporates agreement across experts to mitigate task interference. Based on this formulation, an iterative saliency-aware merging procedure progressively removes non-informative updates while preserving end-to-end connectivity. Furthermore, we extend SA-Merging to introduce rank-wise saliency decomposition for LoRAs without compromising their structural integrity. Extensive experiments on vision and language tasks demonstrate the effectiveness of our saliency-based approach, further reducing the gap between data-free and test-time adaptation methods.

Jungin Park, Jiyoung Lee, Kwanghoon Sohn• 2026

Related benchmarks

TaskDatasetResultRank
Image ClassificationStanford Cars
Accuracy71.8
660
Image ClassificationDTD
Accuracy71
599
Image ClassificationRESISC45
Accuracy86.5
472
Image ClassificationVision Multi-task Suite (SUN397, Cars, RESISC45, EuroSAT, SVHN, GTSRB, MNIST, DTD)
Average Accuracy89.8
104
Natural Language UnderstandingGLUE (test)
QNLI91.5
47
Natural Language UnderstandingGLUE
Average GLUE Score90.2
30
Image ClassificationGTSRB
Accuracy95
21
Image ClassificationMNIST
Accuracy99.6
20
Image Classification8-task vision suite CLIP ViT-L/14 (test)
SUN397 Accuracy82
14
Model MergingLLM Benchmark Family MMLU, TruthfulQA, BBQ, CNN/DailyMail
MMLU Score69.87
5
Showing 10 of 11 rows

Other info

Follow for update