The field of vector graphics generation and understanding is moving towards incorporating advanced reasoning and visual comprehension capabilities into large language models (LLMs) and multimodal models. This is being achieved through the integration of chain-of-thought reasoning, supervised fine-tuning, and reinforcement learning. The use of hybrid reward functions and explicit design rationales is also becoming increasingly prominent, allowing models to generate high-quality vector graphics that are both visually coherent and semantically faithful. Furthermore, there is a growing emphasis on benchmarking and evaluating these models, with a focus on systematic complexity stratification and real-world coverage. Noteworthy papers in this area include Reason-SVG, which pioneers the Drawing-with-Thought paradigm for SVG generation, and SVGenius, which provides a comprehensive benchmark for evaluating LLMs in SVG understanding, editing, and generation. Additionally, papers like ReasonGen-R1 and ControlThinker are making significant contributions to autoregressive image generation and controllable image generation through visual reasoning.