LLM-TTA: Leveraging Large Language Models for Test-Time Adaptation in Image Quality Assessment

Recent test-time adaptation (TTA) methods for image quality assessment (IQA) aim to mitigate the distribution shift between training and testing data by adapting the batch normalization layers of the base IQA model. These methods typically leverage auxiliary tasks—such as group contrastive loss and rank loss—based on feature clustering of low- and high-quality images. However, existing approaches using base models or pre-trained VGG-16 networks struggle to generate meaningful quality-based clusters, limiting their effectiveness. In this work, we introduce LLM-TTA, a novel TTA framework that exploits the perceptual ranking capabilities of pre-trained multi-modal Large Language Models (LLMs). The LLM-TTA algorithm is founded on the observation that, although pre-trained LLMs exhibit limited capability in predicting accurate absolute quality scores, they are highly effective in estimating relative quality rankings and generating more discriminative clusters—both of which are essential for the efficient computation of group contrastive loss. This simple yet impactful insight enables LLM-TTA to substantially improve the adaptability and performance of base IQA models. We validate our method on leading architectures (TReS, MUSIQ, MetaIQA, UIQA) across diverse natural (Koniq-10K, LIVE, PIPAL) and AI-generated (AGIQA-3K, PKU-AGIQA-4K, AIGCIQA2023) image quality assessment datasets. Experimental results demonstrate that LLM-TTA outperforms both existing TTA strategies and standalone LLMs, establishing a new state of the art in test-time IQA adaptation.