ElecVQA-Bench - Binary Defect Assessment (Zero-Shot): leaderboard
ElecVQA-Bench (Zhang et al., 2026) is built from the public InsPLAD UAV power-line inspection imagery; its binary task shows a crop of a power-line asset and asks whether it is normal or defective, with a cannot-judge option always available. Balanced accuracy in percent over 2,021 test crops with abstentions counted as errors, zero-shot with a decision-tree prompt; higher is better.
Metric: Balanced accuracy (%; mean of the defective and normal class recalls on the 2,021-item T2 test split, 766 defective and 1,255 normal crops, with abstentions scored as errors; zero-shot with the P3 prompt (asset description and decision tree) and JSON-constrained guided decoding). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 27B (Non-reasoning) | 66.5 | #199 (Qwen 3.8 27B) |
| 2 | Qwen 2.5 VL 7B | 57.8 | #1796 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=elecvqa-bench-binary-defect-assessment-zero-shot · How It Works · Data refreshed daily, snapshot 2026-11-12.