How Reliable Are Layer Comparisons From Small Probing Datasets

Probing classifiers are used to compare the linguistic information available at different layers of a language model. These comparisons often rely on limited labelled data, although additional training examples can change which layer performs best. We study the accuracy lost when a layer is selected from a small probing dataset and then evaluated with a probe trained on more data. Experiments across encoder and decoder models show that changes in the selected layer are frequent, but many produce little loss in accuracy. Larger losses depend on the checkpoint and, within the examined Qwen3 family, concentrate in the smaller models. In most such cases, the initially selected layer has higher test accuracy with limited data, but another layer benefits more from additional training. Simple geometric scores often make less accurate choices than a fitted linear probe. Controlled experiments also show how feature scaling affects the regularized probe without removing information. Finally, uncertainty estimates for fixed probes do not consistently predict selection errors after further training. The findings support evaluating the accuracy cost of layer selection directly and distinguishing uncertainty about current predictions from changes caused by additional training data.