Rethinking Layer-Wise Information Allocation for Vision Foundation Model Adaptation

Researchers introduced Prompted Information Bottlenecks, a layer-wise prompt-tuning method for frozen vision foundation models.

The approach applies information-bottleneck principles to preserve useful signals and discard irrelevant ones during adaptation.

It improves accuracy and robustness across 34 datasets, including under distribution shift, while updating only a small fraction of parameters.

The work also helps explain why prompt depth and placement matter for performance.