AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment

AutoRubric-T2I introduces an automatic rubric-learning framework for text-to-image reward modeling.

It synthesizes candidate rubrics from preference data, uses VLM judges to apply them, and refines them with sparse logistic regression to keep the most useful rules.

The method achieves strong benchmark performance on MMRB2, TIIF, and UniGenBench++, while making evaluation more transparent and less dependent on human labels.

It also improves downstream text-to-image generation, including diffusion-model training and Flow-GRPO-style optimization.