Recent closed-source multimodal systems have made great advances, but their hidden language for understanding the world remains opaque because of their black-box architectures. In this paper, we use the systems' preference bias to study their hidden language: During the process of compressing the input images (typically containing multiple concepts) into texts and then reconstructing them into images, the systems' inherent preference bias introduces specific shifts in the outputs, disrupting the original input concept co-occurrence. We employ the multi-round "telephone game" to strategically leverage this bias. By observing the co-occurrence frequencies of concepts in telephone games, we quantitatively investigate the concept connection strength in the understanding of multimodal systems, i.e., "hidden language." We also contribute Telescope, a dataset of 10,000+ concept pairs, as the database of our telephone game framework. Our telephone game is test-time scalable: By iteratively running telephone games, we can construct a global map of concept connections in multimodal systems' understanding. Here we can identify preference bias inherited from training, assess generalization capability advancement, and discover more stable pathways for fragile concept connections. Furthermore, we use Reasoning-LLMs to uncover unexpected concept relationships that transcend textual and visual similarities, inferring how multimodal systems understand and simulate the world. This study offers a new perspective on the hidden language of multimodal systems and lays the foundation for future research on the interpretability and controllability of multimodal systems.

Deployed multimodal systems can fail in ways that evaluators did not anticipate.

Abstract--Emergency search and rescue (SAR) operations often require rapid and precise target identification in complex environments where traditional manual drone control is inefficient. This system consists of two aspects: 1) A multimodal system which harnesses the power of Visual Language Model (VLM) and the natural language processing capabilities of ChatGPT-4o (LLM) for scene interpretation. This work aims at improving response times in emergency SAR operations by providing a more intuitive and natural approach to the operator to plan the SAR mission while allowing the drone to carry out that mission in a rapid and safe manner. When tested, our approach was faster on an average by 33.75% when compared with an off-the-shelf autopilot and 54.6% when compared with a human pilot. Search and rescue (SAR) operations in disaster-stricken and hazardous environments require fast and efficient situational assessment to locate survivors and critical infrastructure.

Deployed multimodal models can fail in ways that evaluators did not anticipate. In order to find these failures before deployment, we introduce MultiMon, a system that automatically identifies systematic failures---generalizable, natural-language descriptions that describe categories of individual failures. To uncover systematic failures, MultiMon scrapes for examples of erroneous agreement: inputs that produce the same output, but should not. It then prompts a language model to identify common categories and describe them in natural language. We use MultiMon to find 14 systematic failures (e.g."ignores quantifiers'') of the CLIP text-encoder, each comprising hundreds of distinct inputs (e.g."a shelf with a few/many books''). Because CLIP is the backbone for most state-of-the-art multimodal models, these inputs produce failures in Midjourney 5.1, DALL-E, VideoFusion, and others.

Multimodal fusion is crucial in joint decision-making systems for rendering holistic judgments. Since multimodal data changes in open environments, dynamic fusion has emerged and achieved remarkable progress in numerous applications. However, most existing dynamic multimodal fusion methods lack theoretical guarantees and easily fall into suboptimal problems, yielding unreliability and instability. To address this issue, we propose a Predictive Dynamic Fusion (PDF) framework for multimodal learning. We proceed to reveal the multimodal fusion from a generalization perspective and theoretically derive the predictable Collaborative Belief (Co-Belief) with Mono- and Holo-Confidence, which provably reduces the upper bound of generalization error. Accordingly, we further propose a relative calibration strategy to calibrate the predicted Co-Belief for potential uncertainty. Extensive experiments on multiple benchmarks confirm our superiority. Our code is available at https://github.com/Yinan-Xia/PDF.

Big data contain rich information for machine learning algorithms to utilize when learning important features during classification tasks. Human beings express their emotion using certain words, speech (tone, pitch, speed) or facial expression. Artificial Intelligence approach to emotion classification are largely based on learning from textual information. However, public datasets containing text and speech data provide sufficient resources to train machine learning algorithms for the tack of emotion classification. In this paper, we present novel bimodal deep learning-based architectures enhanced with attention mechanism trained and tested on text and speech data for emotion classification. We report details of different deep learning based architectures and show the performance of each architecture including rigorous error analyses. Our finding suggests that deep learning based architectures trained on different types of data (text and speech) outperform architectures trained only on text or speech. Our proposed attention-based bimodal architecture outperforms several state-of-the-art systems in emotion classification.

In recent years, there has been a significant increase in applications of multimodal signal processing and analysis, largely driven by the increased availability of multimodal datasets and the rapid progress in multimodal learning systems. Well-known examples include autonomous vehicles, audiovisual generative systems, vision-language systems, and so on. Such systems integrate multiple signal modalities: text, speech, images, video, LiDAR, etc., to perform various tasks. A key issue for understanding such systems is the relationship between various modalities and how it impacts task performance. In this paper, we employ the concept of mutual information (MI) to gain insight into this issue. Taking advantage of the recent progress in entropy modeling and estimation, we develop a system called InfoMeter to estimate MI between modalities in a multimodal learning system. We then apply InfoMeter to analyze a multimodal 3D object detection system over a large-scale dataset for autonomous driving. Our experiments on this system suggest that a lower MI between modalities is beneficial for detection accuracy. This new insight may facilitate improvements in the development of future multimodal learning systems.

With recent advancements in Large Multimodal Models (LMMs) across various domains, a novel prompting method called visual referring prompting has emerged, showing significant potential in enhancing human-computer interaction within multimodal systems. This method offers a more natural and flexible approach to human interaction with these systems compared to traditional text descriptions or coordinates. However, the categorization of visual referring prompting remains undefined, and its impact on the performance of LMMs has yet to be formally examined. In this study, we conduct the first comprehensive analysis of LMMs using a variety of visual referring prompting strategies. We introduce a benchmark dataset called VRPTEST, comprising 3 different visual tasks and 2,275 images, spanning diverse combinations of prompt strategies. Using VRPTEST, we conduct a comprehensive evaluation of eight versions of prominent open-source and proprietary foundation models, including two early versions of GPT-4V. We develop an automated assessment framework based on software metamorphic testing techniques to evaluate the accuracy of LMMs without the need for human intervention or manual labeling. We find that the current proprietary models generally outperform the open-source ones, showing an average accuracy improvement of 22.70%; however, there is still potential for improvement. Moreover, our quantitative analysis shows that the choice of prompt strategy significantly affects the accuracy of LMMs, with variations ranging from -17.5% to +7.3%. Further case studies indicate that an appropriate visual referring prompting strategy can improve LMMs' understanding of context and location information, while an unsuitable one might lead to answer rejection. We also provide insights on minimizing the negative impact of visual referring prompting on LMMs.

Generative AI systems (ChatGPT, DALL-E, etc) are expanding into multiple areas of our lives, from art Rombach et al. [2021] to mental health Rob Morris and Kareem Kouddous [2022]; their rapidly growing societal impact opens new opportunities, but also raises ethical concerns. The emerging field of AI alignment aims to make AI systems reflect human values. This paper focuses on evaluating the ethics of multimodal AI systems involving both text and images - a relatively under-explored area, as most alignment work is currently focused on language models. We first create a multimodal ethical database from human feedback on ethicality. Then, using this database, we develop algorithms, including a RoBERTa-large classifier and a multilayer perceptron, to automatically assess the ethicality of system responses.