Researchers have proposed a new framework called MR-IQA-2 for image quality assessment. The system consists of three components: an actor that generates quality reasoning for an input image, an editor that revises the image based on identified quality factors, and a frozen judge that compares the original and edited images to provide reflective supervision. This approach aims to improve faithfulness and reliability in blind IQA by decoupling credit assignment for reasoning and rating. MR-IQA-2 achieves competitive rating alignment with humans across various benchmarks and enables richer visual understanding beyond just rating. --- Why it matters: This matters because current image quality assessment methods often rely on human-provided ratings, which may not accurately reflect the underlying image quality. MR-IQA-2's ability to decouple reasoning and rating supervision can lead to more reliable and faithful image quality assessments, with potential applications in image-quality optimization and related downstream tasks. Source: https://arxiv.org/abs/2608.18579