480 question-answer pairs across 8 open-source hardware projects -- that's the first dataset purpose-built to test whether LLMs actually understand printed circuit board design, and the results are clear: Gemini 3 Flash Preview got 93% accuracy.
The PCB-QA Dataset Turns Schematics into Questions
Researchers manually created PCB-QA, covering component connections, datasheet examination, and SPICE simulation data. Every QA pair comes from real KiCAD-format designs spanning varying complexity. This isn't toy examples -- these are real open-source hardware projects.
JSON Beats Native Formats for LLM Comprehension
The study tested four state-of-the-art models on different PCB representations: graphical PDFs, native KiCAD files, and a proposed JSON-based textual format. Gemini 3 Flash Preview scored 93% with JSON, while other formats lagged. The implication: LLMs need structured text, not images or proprietary file dumps, to make sense of circuit board designs.
What This Means for EDA Workflows
Text-based PCB design formats are finally evaluable by LLMs. The PCB-QA dataset and its JSON representation open the door for integrating LLM assistance into the PCB design lifecycle -- automated checking, question answering, and perhaps even netlist debugging. The paper is the first step, but the 93% accuracy number suggests we're past the proof-of-concept stage.
Source: PCB-QA: Evaluating LLMs over the First Printed Circuit Board Design Question-Answer Dataset
Domain: arxiv.org
Comments load interactively on the live page.