Large language models are now tuned to align with the goals of their creators, namely to be "helpful and harmless." These models should respond helpfully to user questions, but refuse to answer requests that could cause harm. However, adversarial users can construct inputs which circumvent attempts at alignment. In this work, we study adversarial alignment, and ask to what extent these models remain aligned when interacting with an adversarial user who constructs worst-case inputs (adversarial examples). These inputs are designed to cause the model to emit harmful content that would otherwise be prohibited. We show that existing NLP-based optimization attacks are insufficiently powerful to reliably attack aligned text models: even when current NLP-based attacks fail, we can find adversarial inputs with brute force.

Instruction following is crucial in contemporary LLM. However, when extended to multimodal setting, it often suffers from misalignment between specific textual instruction and targeted local region of an image. To achieve more accurate and nuanced multimodal instruction following, we introduce Instruction-guided Visual Masking (IVM), a new versatile visual grounding model that is compatible with diverse multimodal models, such as LMM and robot model. By constructing visual masks for instruction-irrelevant regions, IVM-enhanced multimodal models can effectively focus on task-relevant image regions to better align with complex instructions. Specifically, we design a visual masking data generation pipeline and create an IVM-Mix-1M dataset with 1 million image-instruction pairs. We further introduce a new learning technique, Discriminator Weighted Supervised Learning (DWSL) for preferential IVM training that prioritizes high-quality data samples. Experimental results on generic multimodal tasks such as VQA and embodied robotic control demonstrate the versatility of IVM, which as a plug-and-play tool, significantly boosts the performance of diverse multimodal models, yielding new state-of-the-art results across challenging multimodal benchmarks. Code, model and data are available at https://github.com/2toinf/IVM.

Large-scale multimodal models have shown excellent performance over a series of tasks powered by the large corpus of paired multimodal training data. Generally, they are always assumed to receive modality-complete inputs. However, this simple assumption may not always hold in the real world due to privacy constraints or collection difficulty, where models pretrained on modality-complete data easily demonstrate degraded performance on missing-modality cases. To handle this issue, we refer to prompt learning to adapt large pretrained multimodal models to handle missing-modality scenarios by regarding different missing cases as different types of input. Instead of only prepending independent prompts to the intermediate layers, we present to leverage the correlations between prompts and input features and excavate the relationships between different layers of prompts to carefully design the instructions. We also incorporate the complementary semantics of different modalities to guide the prompting design for each modality. Extensive experiments on three commonly-used datasets consistently demonstrate the superiority of our method compared to the previous approaches upon different missing scenarios. Plentiful ablations are further given to show the generalizability and reliability of our method upon different modality-missing ratios and types.

Web-crawled pretraining datasets underlie the impressive zero-shot evaluation performance of multimodal models, such as CLIP for classification and Stable-Diffusion for image generation. However, it is unclear how meaningful the notion of zero-shot generalization is for such multimodal models, as it is not known to what extent their pretraining datasets encompass the downstream concepts targeted for during zero-shot evaluation. In this work, we ask: How is the performance of multimodal models on downstream concepts influenced by the frequency of these concepts in their pretraining datasets?We comprehensively investigate this question across 34 models and 5 standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting zero-shot generalization, multimodal models require exponentially more data to achieve linear improvements in downstream zero-shot performance, following a sample inefficient log-linear scaling trend. This trend persists even when controlling for sample-level similarity between pretraining and downstream datasets, and testing on purely synthetic data distributions. Furthermore, upon benchmarking models on long-tailed data sampled based on our analysis, we demonstrate that multimodal models across the board perform poorly. We contribute this long-tail test set as the Let it Wag! benchmark to further research in this direction. Taken together, our study reveals an exponential need for training data which implies that the key to zero-shot generalization capabilities under large-scale training data and compute paradigms remains to be found.

Our dataset consists of a novel task suite comprising summarization, perception ( , ), visual reasoning ( , , , , ), and navigation ( , ) tasks. In stark contrast, human experts significantly outperform the state-of-the-art long-context multimodal model, Gemini Pro 1.5 (85.0\% vs. 37.3\%), highlighting a substantial gap in multimodal capabilities.

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. MultiMon can also steer towards failures relevant to specific use cases, such as self-driving cars. We see MultiMon as a step towards evaluation that autonomously explores the long-tail of potential system failures.

The recent explosion of interest in multimodal applications has resulted in a wide selection of datasets and methods for representing and integrating information from different modalities. Despite these empirical advances, there remain fundamental research questions: How can we quantify the interactions that are necessary to solve a multimodal task? Subsequently, what are the most suitable multimodal models to capture these interactions? To answer these questions, we propose an information-theoretic approach to quantify the degree of redundancy, uniqueness, and synergy relating input modalities with an output task. We term these three measures as the PID statistics of a multimodal distribution (or PID for short), and introduce two new estimators for these PID statistics that scale to high-dimensional distributions. To validate PID estimation, we conduct extensive experiments on both synthetic datasets where the PID is known and on large-scale multimodal benchmarks where PID estimations are compared with human annotations. Finally, we demonstrate their usefulness in (1) quantifying interactions within multimodal datasets, (2) quantifying interactions captured by multimodal models, (3) principled approaches for model selection, and (4) three real-world case studies engaging with domain experts in pathology, mood prediction, and robotic perception where our framework helps to recommend strong multimodal models for each application.

Despite surprising performance on zero-shot transfer, pre-training a large-scale multimodal model is often prohibitive as it requires a huge amount of data and computing resources. In this paper, we propose a method (BeamCLIP) that can effectively transfer the representations of a large pre-trained multimodal model (CLIP-ViT) into a small target model (e.g., ResNet-18). For unsupervised transfer, we introduce cross-modal similarity matching (CSM) that enables a student model to learn the representations of a teacher model by matching the relative similarity distribution across text prompt embeddings. To better encode the text prompts, we design context-based prompt augmentation (CPA) that can alleviate the lexical ambiguity of input text prompts. Our experiments show that unsupervised representation transfer of a pre-trained vision-language model enables a small ResNet-18 to achieve a better ImageNet-1K top-1 linear probe accuracy (66.2%) than vision-only self-supervised learning (SSL) methods (e.g., SimCLR: 51.8%, SwAV: 63.7%), while closing the gap with supervised learning (69.8%).