DiagramNet Dataset and Framework for System-Level Diagram Recognition

Researchers have unveiled DiagramNet, the inaugural multimodal dataset tailored for system-level diagrams, which tackles the difficulty of identifying non-standard symbols in chip design schematics. This dataset comprises 10,977 connection annotations and 15,515 chain-of-thought QA pairs, spanning four tasks: Listing, Localization, Connection, and Circuit QA. A progressive training pipeline, alongside a decoupled multi-agent workflow, breaks down visual reasoning into three stages: Perception, Reasoning, and Knowledge. By integrating a 3B-parameter model with this approach, the results exceed those of the 2025 EDA Elite Challenge champion and outperform GPT-5, Claude-Sonnet-4, and Gemini. The findings have been published on arXiv.