Multimodal Large Language Models (MLLMs) have experienced rapid progress in visual recognition tasks in recent years. Given their potential integration into many critical applications, it is important to understand the limitations of their visual perception. In this work, we study whether MLLMs can perceive small visual details as effectively as large ones when answering questions about images. We observe that their performance is very sensitive to the size of the visual subject of the question, and further show that this effect is in fact causal by conducting an intervention study. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then propose training-free visual intervention methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to enhance its perception of small visual details. We evaluate our proposed methods on two widely-used MLLMs and seven visual question answering benchmarks and show that they can significantly improve MLLMs' accuracy without requiring any training. Our results elucidate the risk of applying MLLMs to visual recognition tasks concerning small details and indicate that visual intervention using the model's internal state is a promising direction to mitigate this risk.

Bias issues of neural networks garner significant attention along with its promising advancement. Among various bias issues, mitigating two predominant biases is crucial in advancing fair and trustworthy AI: (1) ensuring neural networks yields even performance across demographic groups, and (2) ensuring algorithmic decision-making does not rely on protected attributes. However, upon the investigation of \pc papers in the relevant literature, we find that there exists a persistent, extensive but under-explored confusion regarding these two types of biases. Furthermore, the confusion has already significantly hampered the clarity of the community and subsequent development of debiasing methodologies. Thus, in this work, we aim to restore clarity by providing two mathematical definitions for these two predominant biases and leveraging these definitions to unify a comprehensive list of papers. Next, we highlight the common phenomena and the possible reasons for the existing confusion. To alleviate the confusion, we provide extensive experiments on synthetic, census, and image datasets, to validate the distinct nature of these biases, distinguish their different real-world manifestations, and evaluate the effectiveness of a comprehensive list of bias assessment metrics in assessing the mitigation of these biases. Further, we compare these two types of biases from multiple dimensions including the underlying causes, debiasing methods, evaluation protocol, prevalent datasets, and future directions. Last, we provide several suggestions aiming to guide researchers engaged in bias-related work to avoid confusion and further enhance clarity in the community.

Recent works have shown that generative models leave traces of their underlying generative process on the generated samples, broadly referred to as fingerprints of a generative model, and have studied their utility in detecting synthetic images from real ones. However, the extend to which these fingerprints can distinguish between various types of synthetic image and help identify the underlying generative process remain under-explored. In particular, the very definition of a fingerprint remains unclear, to our knowledge. To that end, in this work, we formalize the definition of artifact and fingerprint in generative models, propose an algorithm for computing them in practice, and finally study its effectiveness in distinguishing a large array of different generative models. We find that using our proposed definition can significantly improve the performance on the task of identifying the underlying generative process from samples (model attribution) compared to existing methods. Additionally, we study the structure of the fingerprints, and observe that it is very predictive of the effect of different design choices on the generative process.

Multimodal Large Language Models (MLLMs) have recently shown remarkable perceptual capability in answering visual questions, however, little is known about the limits of their perception. In particular, while prior works have provided anecdotal evidence of MLLMs' sensitivity to object size, this phenomenon and its underlying causes have not been explored comprehensively. In this work, we quantitatively study the perception of small visual objects in several state-of-the-art MLLMs and reveal a pervasive limitation in answering questions about small objects in images. Next, we identify four independent factors that can contribute to this limitation -- object quality, size, distractors, and location -- and conduct controlled intervention studies to measure the effect of each factor on MLLMs' perception. In particular, we find that lower object quality and smaller object size can both independently reduce MLLMs' ability to answer visual questions. More surprisingly, we find that the location of the object in the image and the presence of visual distractors can also significantly reduce MLLMs' question answering accuracy. Our study provides a better understanding of the perceptual limitation of MLLMs and contributes new evaluation protocols for analyzing the perception of future MLLMs. To facilitate further investigations, we release our code and data.

Multimodal Large Language Models (MLLMs) have recently achieved promising zero-shot accuracy on visual question answering (VQA) -- a fundamental task affecting various downstream applications and domains. Given the great potential for the broad use of these models, it is important to investigate their limitations in dealing with different image and question properties. In this work, we investigate whether MLLMs can perceive details as well as larger components in images. In particular, we show that their zero-shot accuracy in answering visual questions is very sensitive to the size of the visual subject related to the question, declining up to $45.91\%$ with size. Furthermore, we show that this effect is causal by observing that human visual cropping can significantly mitigate their sensitivity to size. To scale up the usefulness of human cropping, we propose ViCrop, a general framework that utilizes automatic visual cropping to enhance zero-shot VQA of MLLMs. We construct five variants of ViCrop leveraging either external localization models or the decision process of the given MLLM itself. Our results show that ViCrop improves MLLMs' zero-shot accuracy across different VQA datasets, for example, enhances BLIP2-T5's performance by $32.23\%$ on the TextVQA test set. To facilitate further investigation of MLLMs' behaviors, our code is publicly released.

Ensuring a neural network is not relying on protected attributes (e.g., race, sex, age) for predictions is crucial in advancing fair and trustworthy AI. While several promising methods for removing attribute bias in neural networks have been proposed, their limitations remain under-explored. In this work, we mathematically and empirically reveal an important limitation of attribute bias removal methods in presence of strong bias. Specifically, we derive a general non-vacuous information-theoretical upper bound on the performance of any attribute bias removal method in terms of the bias strength. We provide extensive experiments on synthetic, image, and census datasets to verify the theoretical bound and its consequences in practice. Our findings show that existing attribute bias removal methods are effective only when the inherent bias in the dataset is relatively weak, thus cautioning against the use of these methods in smaller datasets where strong attribute bias can occur, and advocating the need for methods that can overcome this limitation.

Ensuring a neural network is not relying on protected attributes (e.g., race, sex, age) for prediction is crucial in advancing fair and trustworthy AI. While several promising methods for removing attribute bias in neural networks have been proposed, their limitations remain under-explored. To that end, in this work, we mathematically and empirically reveal the limitation of existing attribute bias removal methods in presence of strong bias and propose a new method that can mitigate this limitation. Specifically, we first derive a general non-vacuous information-theoretical upper bound on the performance of any attribute bias removal method in terms of the bias strength, revealing that they are effective only when the inherent bias in the dataset is relatively weak. Next, we derive a necessary condition for the existence of any method that can remove attribute bias regardless of the bias strength. Inspired by this condition, we then propose a new method using an adversarial objective that directly filters out protected attributes in the input space while maximally preserving all other attributes, without requiring any specific target label. The proposed method achieves state-of-the-art performance in both strong and moderate bias settings. We provide extensive experiments on synthetic, image, and census datasets, to verify the derived theoretical bound and its consequences in practice, and evaluate the effectiveness of the proposed method in removing strong attribute bias.

Discovering causal relations among semantic factors is an emergent topic in representation learning. Most causal representation learning (CRL) methods are fully supervised, which is impractical due to costly labeling. To resolve this restriction, weakly supervised CRL methods were introduced. To evaluate CRL performance, four existing datasets, Pendulum, Flow, CelebA(BEARD) and CelebA(SMILE), are utilized. However, existing CRL datasets are limited to simple graphs with few generative factors. Thus we propose two new datasets with a larger number of diverse generative factors and more sophisticated causal graphs. In addition, current real datasets, CelebA(BEARD) and CelebA(SMILE), the originally proposed causal graphs are not aligned with the dataset distributions. Thus, we propose modifications to them.

Precision and Recall are two prominent metrics of generative performance, which were proposed to separately measure the fidelity and diversity of generative models. Given their central role in comparing and improving generative models, understanding their limitations are crucially important. To that end, in this work, we identify a critical flaw in the common approximation of these metrics using k-nearest-neighbors, namely, that the very interpretations of fidelity and diversity that are assigned to Precision and Recall can fail in high dimensions, resulting in very misleading conclusions. Specifically, we empirically and theoretically show that as the number of dimensions grows, two model distributions with supports at equal point-wise distance from the support of the real distribution, can have vastly different Precision and Recall regardless of their respective distributions, hence an emergent asymmetry in high dimensions. Based on our theoretical insights, we then provide simple yet effective modifications to these metrics to construct symmetric metrics regardless of the number of dimensions. Finally, we provide experiments on real-world datasets to illustrate that the identified flaw is not merely a pathological case, and that our proposed metrics are effective in alleviating its impact.

Visual Question Answering is a challenging task, as it requires seamless interaction between perceptual, linguistic, and background knowledge systems. While the recent progress of visual and natural language models like BLIP has led to improved performance on this task, we lack understanding of the ability of such models to perform on different kinds of questions and reasoning types. As our initial analysis of BLIP-family models revealed difficulty with answering fine-detail questions, we investigate the following question: Can visual cropping be employed to improve the performance of state-of-the-art visual question answering models on fine-detail questions? Given the recent success of the BLIP-family models, we study a zero-shot and a fine-tuned BLIP model. We define three controlled subsets of the popular VQA-v2 benchmark to measure whether cropping can help model performance. Besides human cropping, we devise two automatic cropping strategies based on multi-modal embedding by CLIP and BLIP visual QA model gradients. Our experiments demonstrate that the performance of BLIP model variants can be significantly improved through human cropping, and automatic cropping methods can produce comparable benefits. A deeper dive into our findings indicates that the performance enhancement is more pronounced in zero-shot models than in fine-tuned models and more salient with smaller bounding boxes than larger ones. We perform case studies to connect quantitative differences with qualitative observations across question types and datasets. Finally, we see that the cropping enhancement is robust, as we gain an improvement of 4.59% (absolute) in the general VQA-random task by simply inputting a concatenation of the original and gradient-based cropped images. We make our code available to facilitate further innovation on visual cropping methods for question answering.