PCB-Bench (digailab)

PCB-Bench is a benchmark from ICLR 2026 for evaluating large language models on PCB placement and routing. It spans three task settings: text-based question answering, multimodal image-and-text reasoning over PCB layout images, and design understanding where the model describes a full board from an EDA editor screenshot. The README states all models are evaluated under a unified zero-shot setting.

The text setting holds about 1,800 expert-written free-form questions, each with a single-choice version, for a total of about 3,700 questions covering component placement, routing strategies, and design rule compliance. Topics include signal integrity, EMI/EMC, power planning, differential pairs, and DFM. The multimodal setting adds about 500 problems that need joint reading of layout images and technical prompts, including component identification, trace reasoning, and differential-pair continuity.

The design-understanding setting uses 174 complete real-world PCB projects from OSHWHub, each with schematics, placement and routing files, design descriptions, component libraries, and EDA screenshots. Choice questions are scored by top-1 accuracy, free-form answers by BERTScore and Sentence-BERT similarity, and the design-understanding task additionally by precision, recall, and F1-score. The README states the benchmark ships with evaluation scripts and configuration files for reproduction.