Multimodal large language models (MLLMs) are increasingly used to evaluate text-to-image (TTI) generation systems, providing automated judgments based on visual and textual context. However, these "judge" models often suffer from biases, overconfidence, and inconsistent performance across diverse image domains. While prompt ensembling has shown promise for mitigating these issues in unimodal, text-only settings, our experiments reveal that standard ensembling methods fail to generalize effectively for TTI tasks. To address these limitations, we propose a new multimodal-aware method called Multimodal Mixture-of-Bayesian Prompt Ensembles (MMB). Our method uses a Bayesian prompt ensemble approach augmented by image clustering, allowing the judge to dynamically assign prompt weights based on the visual characteristics of each sample. We show that MMB improves accuracy in pairwise preference judgments and greatly enhances calibration, making it easier to gauge the judge’s true uncertainty. In evaluations on two TTI benchmarks, HPSv2 and MJBench, MMB outperforms existing baselines in alignment with human annotations and calibration across varied image content. Our findings highlight the importance of multimodal-specific strategies for judge calibration and suggest a promising path forward for reliable large-scale TTI evaluation.

Recent dataset deduplication techniques have demonstrated that content-aware dataset pruning can dramatically reduce the cost of training Vision-Language Pretrained (VLP) models without significant performance losses compared to training on the original dataset. These results have been based on pruning commonly used image-caption datasets collected from the web – datasets that are known to harbor harmful social biases that may then be codified in trained models. In this work, we evaluate how deduplication affects the prevalence of these biases in the resulting trained models and introduce an easy-to-implement modification to the recent SemDeDup algorithm that can reduce the negative effects that we observe. When examining CLIP-style models trained on deduplicated variants of LAION-400M, we find our proposed FairDeDup algorithm consistently leads to improved fairness metrics over SemDeDup on the FairFace and FACET datasets while maintaining zero-shot performance on CLIP benchmarks.

Recent work in vision-and-language demonstrates that large-scale pretraining can learn generalizable models that are efficiently transferable to downstream tasks. While this may improve dataset-scale aggregate metrics, analyzing performance around hand-crafted subgroups targeting specific bias dimensions reveals systemic undesirable behaviors. However, this subgroup analysis is frequently stalled by annotation efforts, which require extensive time and resources to collect the necessary data. Prior art attempts to automatically discover subgroups to circumvent these constraints but typically leverages model behavior on existing task-specific annotations and rapidly degrades on more complex inputs beyond "tabular" data, none of which study vision-and-language models. This paper presents VLSlice, an interactive system enabling user-guided discovery of coherent representation-level subgroups with consistent visiolinguistic behavior, denoted as vision-and-language slices, from unlabeled image sets. We show that VLSlice enables users to quickly generate diverse high-coherency slices in a user study (n=22) and release the tool publicly.