Trustworthy Evaluation of Vision–Language Models for UAV Power-Line Inspection: A Resolution- and Parity-Controlled Benchmark

Vision–language models (VLMs) are often reported to outperform vision-only backbones for UAV power-line inspection. We examine that claim using ElecVQA-Bench, a 56,972-item resource derived from InsPLAD, and an audit of partition, evaluated items, label space, replication, input resolution, and side information. The original split has substantial source-photograph dependence and is used for paired diagnostic comparisons. Reported tower-disjoint retraining provides a limited check on new-tower performance within the same source dataset. On the original binary task, no difference is detected by a paired item-level test between the selected Swin and Qwen runs; this does not establish equivalence. Fine-grained results change with resolution, prompt information, class support, and seed. A 14-tower, three-seed comparison changes direction across seeds, and classes represented by only a few items substantially affect macro recall. The contribution is a reproducible task representation and reporting protocol for examining these dependencies, rather than a general ranking of model families or evidence of end-to-end UAV deployment.