When Does Representation Stability Predict Retrieval Robustness? A Within- and Across-Architecture Analysis Under Resolution Degradation

Representation stability, how much an image embedding shifts under resolution degradation,is an intuitive candidate for anticipating downstream robustness in cosine-similarity retrieval,but whether it predicts retrieval success has not been tested directly. This study examines therelationship within a given architecture (does an image’s stability predict its own retrievalsuccess) and across architectures (does the more stable model retrieve more accurately). Threepretrained architectures (ResNet18, VGG16, ViT-B/16) were evaluated on ChestMNIST andSTL-10 under progressive resolution degradation, with stability as cosine similarity to base-resolution embeddings and retrieval as Top-1 accuracy against a base-resolution gallery.Stability was tested with two-way repeated-measures analysis of variance (ANOVA);generalized estimating equations tested whether stability predicted retrieval, supplemented bycorrelation, collinearity, and mediation analyses; McNemar’s test compared architectures.Within each architecture, greater stability was strongly associated with successful retrievalacross all six model–dataset combinations (odds ratios 15.47–642.14 per standard deviation(SD) increase). Across architectures, stability ranking did not reliably predict accuracy ranking:ResNet18 was most stable throughout, but ViT-B/16 outperformed it on STL-10 below 40 px,with no reversal on ChestMNIST. These results confirm the assumption within a fixedarchitecture but not universally: stability and retrieval are related but not interchangeable.