Hidden Clones: Exposing and Fixing Family Bias in Vision-Language Model Ensembles

Vision-Language Models (VLMs) such as CLIP are now foundational to multimodal systems, yet their robustness to spurious correlations remains poorly understood at scale. We present the first large-scale empirical study of 194 publicly available VLMs, including 16 model families, covering a wide range of model sizes, 24 training datasets, and three evaluation benchmarks, namely ImageNet (overall performance), CelebA (typical single-attribute bias), and UrbanCars (complex multi-attribute biases). Across these settings, the Spearman correlation between model scale and performance weakens as evaluation shifts from ImageNet ($ρ{=}0.68$) to single-attribute ($ρ{=}0.48$) and further to multi-attribute ($ρ{=}0.05$) bias benchmarks. In contrast, properties of the training data (size and quality) show more consistent relationships with worst-group accuracy across both bias benchmarks. Notably, curated datasets yield improvements of up to 25% over uncurated alternatives at a comparable scale. Finally, the effect of architectural choices (e.g., patch size, image resolution) is highly context-dependent, varying with the nature of the benchmark, including the type of bias and its spatial distribution within images.

With the growing number and diversity of Vision-Language Models (VLMs), many works explore language-based ensemble, collaboration, and routing techniques across multiple VLMs to improve multi-model reasoning. In contrast, we address the diverse model selection using both vision and language modalities. We introduce focal error diversity to capture complementary reasoning across VLMs and a CKA-based focal diversity metric (CKA-focal) to measure disagreement in their visual embeddings. On the constructed ensemble surface from a pool of candidate VLMs, we applied a Genetic Algorithm to effectively prune out those component VLMs that do not add value to the fusion performance. We identify the best combination for each task as well as fuse the outputs of each VLMs in the model pool, and show that heterogeneous models can capture epistemic uncertainty dynamically and mitigate hallucinations. Our V3Fusion approach is capable of producing dual focal-diversity fused predictions with high performance for vision-language reasoning, even when there is no majority consensus or the majority of VLMs make incorrect predictions. Extensive experiments validate V3Fusion on four popular VLM benchmarks (A-OKVQA, MMMU, MMMU-Pro, and OCR-VQA). The results show that V3Fusion outperforms the best-performing VLM on MMMU by 8.09% and MMMU-Pro by 4.87% gain in accuracy. For generative tasks, V3Fusion outperforms Intern-VL2-8b and Qwen2.5-VL-7b, the top-2 VLM performers on both A-OKVQA and OCR-VQA. Our code and datasets are available at https://github.com/sftekin/v3fusion.

The reliability of VLM-as-a-Judge is critical for the automatic evaluation of vision-language models (VLMs). Despite recent progress, our analysis reveals that VLM-as-a-Judge often pays limited attention to the image when making decisions. Instead, they often blindly favor the more informative answer, even when they can recognize it conflicts with the image content. We call this problem informativeness bias, which significantly undermines judge reliability. To address it, we propose BIRCH (Balanced Informativeness and CoRrectness with a Truthful AnCHor), a judging paradigm that first corrects inconsistencies with the image content in candidate answers, and then compares the answers against this corrected version. This shifts the judge's focus from informativeness to image-grounded correctness. Experiments on multiple models and benchmarks show that BIRCH reduces informativeness bias by up to 17%, resulting in performance gains of up to 9.8%. Our work reveals an overlooked but fundamental flaw in current VLM-as-a-Judge systems and highlights the need for more principled designs.