Spurious bias, a tendency to exploit spurious correlations between superficial input attributes and prediction targets, has revealed a severe robustness pitfall in classical machine learning problems. Multimodal Large Language Models (MLLMs), which leverage pretrained vision and language models, have recently demonstrated strong capability in joint vision-language understanding. However, both the presence and severity of spurious biases in MLLMs remain poorly understood. In this work, we address this gap by analyzing the spurious biases in the multimodal setting and uncovering the specific inference-time data patterns that can manifest this problem. To support this analysis, we introduce MM-SpuBench, a comprehensive, human-verified benchmark dataset consisting of image-class pairs annotated with core and spurious attributes, grounded in our taxonomy of nine distinct types of spurious correlations. The benchmark is constructed using human-interpretable attribute information to capture a wide range of spurious patterns reflective of real-world knowledge. Leveraging this benchmark, we conduct a comprehensive evaluation of the state-of-the-art open-source and proprietary MLLMs with both standard accuracy and the proposed Conditional Generation Likelihood Advantage (CGLA). Our findings highlight the persistence of reliance on spurious correlations and the difficulty of mitigation on our benchmark. We hope this work can inspire new technical strides to mitigate these biases. Our benchmark is publicly available at https://huggingface.co/datasets/mmbench/MM-SpuBench.

Deep neural networks often learn and rely on spurious correlations, i.e., superficial associations between non-causal features and the targets. For instance, an image classifier may identify camels based on the desert backgrounds. While it can yield high overall accuracy during training, it degrades generalization on more diverse scenarios where such correlations do not hold. This problem poses significant challenges for out-of-distribution robustness and trustworthiness. Existing methods typically mitigate this issue by using external group annotations or auxiliary deterministic models to learn unbiased representations. However, such information is costly to obtain, and deterministic models may fail to capture the full spectrum of biases learned by the models. To address these limitations, we propose Evidential Alignment, a novel framework that leverages uncertainty quantification to understand the behavior of the biased models without requiring group annotations. By quantifying the evidence of model prediction with second-order risk minimization and calibrating the biased models with the proposed evidential calibration technique, Evidential Alignment identifies and suppresses spurious correlations while preserving core features. We theoretically justify the effectiveness of our method as capable of learning the patterns of biased models and debiasing the model without requiring any spurious correlation annotations. Empirical results demonstrate that our method significantly improves group robustness across diverse architectures and data modalities, providing a scalable and principled solution to spurious correlations.

Spurious correlations are brittle associations between certain attributes of inputs and target variables, such as the correlation between an image background and an object class. Deep image classifiers often leverage them for predictions, leading to poor generalization on the data where the correlations do not hold. Mitigating the impact of spurious correlations is crucial towards robust model generalization, but it often requires annotations of the spurious correlations in data -- a strong assumption in practice. In this paper, we propose a novel learning framework based on meta-learning, termed SPUME -- SPUriousness-aware MEta-learning, to train an image classifier to be robust to spurious correlations. We design the framework to iteratively detect and mitigate the spurious correlations that the classifier excessively relies on for predictions. To achieve this, we first propose to utilize a pre-trained vision-language model to extract text-format attributes from images. These attributes enable us to curate data with various class-attribute correlations, and we formulate a novel metric to measure the degree of these correlations' spuriousness. Then, to mitigate the reliance on spurious correlations, we propose a meta-learning strategy in which the support (training) sets and query (test) sets in tasks are curated with different spurious correlations that have high degrees of spuriousness. By meta-training the classifier on these spuriousness-aware meta-learning tasks, our classifier can learn to be invariant to the spurious correlations. We demonstrate that our method is robust to spurious correlations without knowing them a priori and achieves the best on five benchmark datasets with different robustness measures. Our code is available at https://github.com/gtzheng/SPUME.

Meta-learning has shown great promise in few-shot image classification where only a small amount of labeled data is available in each classification task. Many training tasks are provided to train a meta-model that can quickly learn new and similar concepts with few labeled samples. Data augmentation is often used to augment training tasks to avoid overfitting. However, existing data augmentation methods are often manually designed and fixed during training, ignoring training dynamics and the difference between various meta-learning settings specified by meta-model architectures and meta-learning algorithms. To address this problem, we add a task transformation layer between a training task and a meta-model such that the right amount of perturbation is added to training tasks for a certain meta-learning setting at a certain training stage. By jointly optimizing the task transformation layer and the meta-model, we avoid the risk of providing tasks that are either too easy or too difficult during training. We design the task transformation layer as a stochastic transformation function, adding the flexibility in how a training task can be transformed. We leverage differentiable data augmentations as the building blocks of the task transformation function for efficient optimization. Extensive experiments show that our method can consistently improve the few-shot generalization performance of various meta-models trained with different meta-learning algorithms, meta-model architectures, and datasets.

In few-shot image classification, it is challenging for deep neural networks to infer the true class of an image when it contains multiple class-unrelated objects and its label carries no semantic meanings, such as Class 1 or Class 2. In contrast, knowing what are not important in a typical classification task, humans can quickly identify the right class objects with very few images. In this paper, we propose to extract semantic features from a given dataset to filter out class-unrelated objects in a few-shot task. Each semantic feature is meta-learned and represents a common object or pattern shared by many tasks. The strengths of these features in a given image are adjusted by an importance kernel encoding the meta-learned knowledge such that class-unrelated objects can be suppressed, and the few-shot classification performance can be improved. To facilitate learning and identifying semantic features in an image, we further propose an image representation decomposition module to decouple complex correlations between objects in an image embedding. The experimental analysis demonstrates the effectiveness of our method, especially in the extremely low-shot cases.