Real-world engineering drawings are often large-format and packed with information, requiring models to combine fine-grained reading with complex reasoning across multiple views. Existing benchmarks tend to focus on local regions or individual views, leaving a gap in systematic evaluation of realistic, large-format drawings. To address this, the authors introduce LED-BENCH, a multidisciplinary benchmark covering mechanical, architectural, and transportation engineering drawings.

The benchmark was used to evaluate 11 representative multimodal large language models, including 8 proprietary and 3 open-source models. Results show substantial room for improvement: the best performer, GPT-5.6 Sol, reached only 52.65% average accuracy. Accuracy also varied across engineering domains, and all models scored lower on questions that demanded broader evidence integration across different parts of a drawing.

The findings highlight persistent challenges for current multimodal models in fine-grained understanding, evidence integration, and complex reasoning over large-format engineering drawings. Since this is a single source, there are no conflicting results to compare; the paper's own conclusions are reported here as presented by the authors.