CDABench: A Benchmark for Grounded Generation of Component Design Assets

Generating an EDA component library requires accurate schematic symbols, PCB footprints, and 3D package models, correctly bound to orderable part num bers . Family datasheets complicate this task through scattered specifications and shared assets across variants . We introduce CDABench, a benchmark of di verse datasheets paired with official or engineer-verified reference libraries . Its five tasks evaluate asset generation from verified evidence, relation graph recov ery from datasheets, and end-to-end library generation, accounting for missing and extra elements, asset quality, and binding correctness . We also present a pipeline combining evidence retrieval, graph-guided refinement, and a format independent domain-specific language (DSL) with construction, editing, and ex ecutable checks . Across large language models, entity and binding recovery is more reliable than accurate file generation . Footprint and 3D generation remain challenging even with verified inputs, highlighting difficulties in translating spec ifications into precise geometry. Ablations show that graph guidance improves relation recovery and that the DSL improves symbol and footprint quality over template-based and native-format or coordinate-based generation .