Physical World & Real-World ImpactXi Jiang, Jian Li, Chengjie Wang et al.
Xi Jiang et al. release the MMAD benchmark
MMAD benchmark released — 8,366 industrial images and 39,672 questions. The best performer measured by the authors, GPT-4o, averaged 74.9% accuracy, and the paper states that this falls far short of industrial requirements
Why it matters
The upper bound for the idea of handing defect judgment directly to general-purpose multimodal LLMs. 74.9% is not a level usable for final-inspection decisions
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2410.09453Added in Week 35, 2026 · See that week