In this paper, we propose a new benchmark, Generative DyTAG Benchmark (GDGB), for dynamic text attribute graph (DyTAG) generation. To improve the low text quality and discriminative task-centered research of existing DyTAG datasets, we construct eight DyTAG datasets with high-quality text attributes and define two new generation tasks, Transductive Dynamic Graph Generation (TDGG) and Inductive Dynamic Graph Generation (IDGG). TDGG generates DyTAG based on a given set of source and destination nodes, while IDGG models dynamic graph expansion including new node creation. We present multifaceted metrics that evaluate structural, temporal, and textual qualities, and GAG-General, an LLM-based multi-agent generation framework for DyTAG generation, to enable rigorous evaluation. Experimental results show that GDGB enables rigorous evaluation of TDGG and IDGG, and demonstrates the interaction of structural and textual features in DyTAG generation.
