Few-shot adaptation in vision-language models typically depends on the encoder's final global representation, leaving complementary information in intermediate layers underused. According to a new arXiv paper, this common design choice may limit performance when only a few examples are available.
The paper introduces LAMP (Layerwise Adaptive Mixing of Predictors), a training-free approach that adaptively mixes predictors from different layers. The authors note that combining these features is challenging because their dimensions can differ, and LAMP is designed to address that difficulty without requiring fine-tuning.
The abstract does not include experimental results or implementation details, so the practical gains remain unspecified. Still, the proposal points to a simple direction for squeezing more signal out of pretrained encoders in few-shot settings.