RULER introduces instance-aware rubric rewards for training SVG-generation models with reinforcement learning, replacing scalar image-quality metrics that transfer poorly to vector graphics and are prone to reward hacking. Each instruction becomes a six-item rubric spanning semantic, visual, and stylistic axes; a vision-language judge scores rendered rollouts against it, and the weighted result is optimized via Group Relative Policy Optimization without any paired ground-truth SVGs. On MMSVG-Illustration and MMSVG-Icon, RULER lifts rubric scores substantially, matching a much larger baseline model.
