LADBench: A Benchmark for Logical Fault Detection in Images

A new benchmark called LAD-bench reveals that leading vision-language models struggle to detect logical anomalies in images, achieving only 70.11% overall accuracy even with progressive hints, according to a paper posted to arXiv on June 16, 2026 [1][2]. The benchmark, introduced by Sahasra Kondapalli and collaborators, comprises more than 1,000 curated synthetic images containing logical faults across four domains: Residential, Urban, Collaborative, and Nature [1][2]. Unlike existing anomaly benchmarks that emphasize visual errors or direct prompting, LAD-bench tests a model's capacity for the physical and social common sense required for open-world deployment [2]. The paper proposes a Tiered Prompting Protocol based on progressive disclosure, which measures how much explicit assistance a model needs to localize and reason about a logical fault [2]. Evaluations of leading foundation models show that implicit logical fault detection remains unsolved [2]. The best-performing model reached only 70.11% overall accuracy [1][2]. The authors note that models often fail to identify anomalies even after receiving explicit hints in deeper tiers of the prompting protocol [2]. By surfacing these limitations in sequential multimodal reasoning, the benchmark offers a framework for advancing the safety, reliability, and cognitive alignment of autonomous visual systems [2]. The paper was submitted to arXiv, an open-access repository of electronic preprints that has hosted scientific papers since 1991 and now receives about 24,000 submissions per month [6]. The dataset and code for LAD-bench are available on Hugging Face [2].