Fooling the LVLM Judges: Visual Biases in LVLM-Based Evaluation
This work exposes critical vulnerabilities in LVLM-based evaluation systems, highlighting an urgent need for more robust judges to ensure fair assessments in AI and vision-language tasks.
The study investigated whether adversarial visual manipulations can systematically deceive large vision-language model (LVLM) judges into assigning unfairly high scores for text-image alignment, revealing that all tested LVLM judges were vulnerable across domains, consistently inflating scores for manipulated images.
Recently, large vision-language models (LVLMs) have emerged as the preferred tools for judging text-image alignment, yet their robustness along the visual modality remains underexplored. This work is the first study to address a key research question: Can adversarial visual manipulations systematically fool LVLM judges into assigning unfairly inflated scores? We define potential image induced biases within the context of T2I evaluation and examine how these biases affect the evaluations of LVLM judges. Moreover, we introduce a novel, fine-grained, multi-domain meta-evaluation benchmark named FRAME, which is deliberately constructed to exhibit diverse score distributions. By introducing the defined biases into the benchmark, we reveal that all tested LVLM judges exhibit vulnerability across all domains, consistently inflating scores for manipulated images. Further analysis reveals that combining multiple biases amplifies their effects, and pairwise evaluations are similarly susceptible. Moreover, we observe that visual biases persist under prompt-based mitigation strategies, highlighting the vulnerability of current LVLM evaluation systems and underscoring the urgent need for more robust LVLM judges.