Abstract
Vision-language models (VLMs) offer a promising approach to estimate floodwater depth from images, a critical metric for emergency response. While their accuracy has been benchmarked against human estimators, their reasoning has mostly been assessed qualitatively and post hoc. In this work, we hypothesize that bounding boxes provide a systematic framework to evaluate VLM reasoning and localization. To test this, we inferenced the GLM-4.5V VLM on a dataset of flooding images and compared its depth estimates and reference-object bounding boxes against the consensus of four human annotators under four perturbation scenarios. We employ the distance intersection over union (DIoU) metric to quantify the agreement between bounding boxes, and the mean absolute error (MAE) to characterize the accuracy of floodwater depth estimates. Results show that VLMs use bounding boxes with accuracy comparable to textual descriptions, though below human performance, and that numerical estimates depend on the visual content of the chosen boxes. This demonstrates that bounding boxes can capture VLM reasoning traces and systematically reveal failure cases previously identified only through qualitative analysis.