📄 中文摘要
该研究提出了一种逐步生成矢量草图的方法。通过使用一种新颖的多回合过程奖励强化学习方法，结合监督微调，训练了基于多模态语言模型的智能体。研究中引入了一个名为ControlSketch-Part的新数据集，包含丰富的草图部件级注释，利用一种新颖的通用自动注释管道将矢量草图分割为语义部分，并通过结构化的多阶段标注过程为部分分配路径。结果表明，结合结构化的部件级数据并通过过程提供视觉反馈，使得生成的文本到矢量草图具备可解释性、可控性和局部可编辑性。
逐步教导智能体进行草图绘制
出处: Teaching an Agent to Sketch One Part at a Time
发布: 2026年3月23日
📄 English Summary
Teaching an Agent to Sketch One Part at a Time
This research presents a method for generating vector sketches one part at a time. A multi-modal language model-based agent is trained using a novel multi-turn process-reward reinforcement learning approach following supervised fine-tuning. The study introduces a new dataset called ControlSketch-Part, which contains rich part-level annotations for sketches, obtained through a novel, generic automatic annotation pipeline that segments vector sketches into semantic parts and assigns paths to these parts via a structured multi-stage labeling process. Results indicate that incorporating structured part-level data and providing the agent with visual feedback throughout the process enables interpretable, controllable, and locally editable text-to-vector sketch generation.