Beyond Scalar Rewards by Internalizing Reasoning into Score Distributions
Researchers have developed a new framework called Z-Reward for improving text-to-image generation models. This system uses a teacher-student approach where a large vision-language model (VLM) acts as the teacher, inferring score distributions based on reasoning. A smaller student VLM is then trained to mimic these distributions, enabling efficient reward deployment without requiring explicit reasoning during inference. The Z-Reward framework demonstrated significant improvements in human preference accuracy compared to existing methods and enhanced text-to-image optimization. AI
IMPACT Introduces a novel reward modeling technique that could enhance the quality and controllability of text-to-image generation models.