RULER：面向 SVG 生成的实例感知评分标准奖励

AI 摘要

该论文提出 RULER 方法，用于从自然语言指令生成 SVG 代码的强化学习。作者发现用多轴评分标准（rubric）提示视觉语言模型评判，比 CLIP、Aesthetic 等标量指标更符合人类判断，且能避免奖励黑客问题。RULER 将每条指令转化为包含语义、视觉、风格六个条目的实例感知评分标准，由评判 VLM 逐项打分并加权作为奖励，通过 GRPO 优化，无需配对 SVG 真值或人类偏好标签。在 MMSVG-Illustration 和 MMSVG-Icon 上，评分从 0.432/0.395 提升至 0.693/0.683，超过专用 SVG 模型并匹敌更大的 DeepSeek-V3。

正文节选

RULER: Instance-aware Rubric Rewards for SVG Generation Abstract Generating Scalable Vector Graphics (SVG) code from natural-language instructions is an open-ended task without absolute visual ground truth, leaving both evaluation and policy optimization without a faithful signal. Scalar metrics (CLIP, Aesthetic) calibrated on natural images transfer poorly to stylized vector content, and reusing them as RL rewards triggers reward hacking. We address both limitations with rubric-based scoring. W

发布时间：—

抓取时间：2026-09-23 11:10

来源机构：Hugging Face