AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment
Researchers have developed AutoRubric-T2I, a novel framework for text-to-image generation that automatically creates and refines explicit rubrics. These rubrics guide Vision-Language Models (VLMs) in evaluating image quality and prompt alignment, significantly reducing the need for extensive human preference data. The system synthesizes reasoning traces into candidate rules and uses a logistic regression refiner to select the most discriminative ones, achieving high-quality, interpretable reward signals with minimal annotation. AI
IMPACT Enables more efficient and interpretable reward modeling for text-to-image generation, reducing data annotation costs.