Vision encoder layers with certain weight traits improve fine-tuning results

Pre-PEFT Probing: Weight Statistics and Perturbation Robustness for Layer Selection in VLM Vision Encoders

Computer Vision and Pattern RecognitionArtificial Intelligence

Summary

Fine-tuning large vision-language computer models usually involves adjusting many parts at once, which can be inefficient. The authors found a way to check specific layers in these models before fine-tuning by looking at their internal weight properties and stability to small changes. These checks can predict which layers will improve performance the most when fine-tuned individually. This approach helps reduce the amount of training needed while still getting good results.

What this means in practice

Authors

Qingtao Xia, Jiahua Bao, Siyao Cheng, Jie Liu

Abstract

We propose a pre-fine-tuning probing method for Parameter-Efficient Fine-Tuning (PEFT) layer selection, aiming to obtain more stable and higher gains with fewer trainable parameters when adapting large vision--language models (VLMs). Unlike the common practice of applying LoRA and other adapters to all layers at once---where layer selection often relies on heuristic rules---we focus on the vision encoder and directly evaluate the "adaptability'' of each Transformer layer. Specifically, we characterize each layer from two perspectives: (i) the statistical properties of its Q/K/V projection weights (e.g., norms and condition numbers); (ii) robustness under controlled parameter perturbations. We then systematically compare these indicators with the downstream performance gains brought by applying PEFT to a single layer. Across experiments covering seven benchmarks and five PEFT variants, we observe a consistent correlation: layers (or matrices) with larger weight norms and higher condition numbers are usually more robust to perturbations and are more likely to yield larger fine-tuning gains. These results show that distribution-statistics analysis and perturbation tests before fine-tuning can provide practical signals for adaptation-layer selection, thereby maintaining or improving performance while reducing trainable parameters.