Knossos and Ariadne: Benchmarking and Learning Complete Diagram Topology Extraction with Vision-Language Models
The paper introduces a 19,200-diagram benchmark and a structured VLM method for extracting full diagram graphs.
Knossos covers six domains and provides exact annotations for nodes, edges, relation types, and connector geometry. Ariadne splits the problem into node inventory extraction and source-conditioned edge prediction. In the reported experiments, training on Knossos improved smaller open-source VLMs, and Ariadne produced the highest average Edge F1 among the evaluated methods on the benchmark. The authors say both backbone variants also improved over unadapted versions on a real-world external benchmark. ArXiv · AI/CL/LG's note
Knossos covers six domains and provides exact annotations for nodes, edges, relation types, and connector geometry. Ariadne splits the problem into node inventory extraction and source-conditioned edge prediction. In the reported experiments, training on Knossos improved smaller open-source VLMs, and Ariadne produced the highest average Edge F1 among the evaluated methods on the benchmark. The authors say both backbone variants also improved over unadapted versions on a real-world external benchmark. ArXiv · AI/CL/LG's note
score 5