PCB-Bench: Benchmarking LLMs for PCB Placement and Routing (ICLR 2026) (github.com)

PCB-Bench, set to debut at ICLR 2026, introduces the first comprehensive benchmark for evaluating multimodal large language models (LLMs) specifically in the realm of printed circuit board (PCB) placement and routing. Addressing the existing gap in standardized benchmarks for PCB engineering, PCB-Bench integrates textual, visual, and design artifact data into a unified framework. It encompasses three task settings: free-form and multiple-choice questions on PCB knowledge, multimodal Q&A based on layout images, and descriptions generated from EDA tool screenshots of PCB designs. This benchmark is significant for the AI/ML community because it offers a systematic methodology for assessing LLM performance in a critical engineering domain, facilitating advancements in both AI interpretability and design automation. By including approximately 3,700 generated questions from real-world PCB projects—with supporting artifacts accessible from OSHWHub—it allows researchers to examine model responses across various contexts. The benchmark supports reproducibility with open licensing, making it a valuable resource for researchers aiming for meaningful comparisons in AI-driven PCB design, and represents a vital step towards integrating advanced AI technologies into complex engineering tasks.