DUDA-Bench benchmarks agents on urban diagnosis tasks

DUDA-Bench is a new hierarchical benchmark for evaluating LLM agents on multimodal urban diagnosis tasks, covering 12 cities and 106 tasks. The benchmark reveals a significant gap between an agent's ability to perform isolated analytical tasks and its ability to complete end-to-end diagnostic workflows.