Sparse Attention Vectors: Generative Multimodal Model Features Are Discriminative Vision-Language Classifiers

Generative Large Multimodal Models (LMMs) like LLaVA and Qwen-VL excel at a wide variety of vision-language (VL) tasks such as image captioning or visual question answering. Despite strong performance, LMMs are not directly suited for foundational discriminative vision-language tasks (i.e., tasks requiring discrete label predictions) such as image classification and multiple-choice VQA. One key challenge in utilizing LMMs for discriminative tasks is the extraction of useful features from generative models. To overcome this issue, we propose an approach for finding features in the model's latent space to more effectively leverage LMMs for discriminative tasks. Toward this end, we present Sparse Attention Vectors (SAVs) -- a finetuning-free method that leverages sparse attention head activations (fewer than 1\% of the heads) in LMMs as strong features for VL tasks. With only few-shot examples, SAVs demonstrate state-of-the-art performance compared to a variety of few-shot and finetuned baselines on a collection of discriminative tasks. Our experiments also imply that SAVs can scale in performance with additional examples and generalize to similar tasks, establishing SAVs as both effective and robust multimodal feature representations.

NaturalBench: Evaluating Vision-Language Models on Natural Adversarial Samples

Vision-language models (VLMs) have made significant progress in recent visual-question-answering (VQA) benchmarks that evaluate complex visio-linguistic reasoning. However, are these models truly effective? In this work, we show that VLMs still struggle with natural images and questions that humans can easily answer, which we term natural adversarial samples. We also find it surprisingly easy to generate these VQA samples from natural image-text corpora using off-the-shelf models like CLIP and ChatGPT. We propose a semi-automated approach to collect a new benchmark, NaturalBench, for reliably evaluating VLMs with 10,000 human-verified VQA samples. Crucially, we adopt a $\textbf{vision-centric}$ design by pairing each question with two images that yield different answers, preventing blind solutions from answering without using the images. This makes NaturalBench more challenging than previous benchmarks that can be solved with commonsense priors. We evaluate 53 state-of-the-art VLMs on NaturalBench, showing that models like LLaVA-OneVision, Cambrian-1, Llama3.2-Vision, Molmo, Qwen2-VL, and even GPT-4o lag 50%-70% behind human performance (over 90%). We analyze why NaturalBench is hard from two angles: (1) Compositionality: Solving NaturalBench requires diverse visio-linguistic skills, including understanding attribute bindings, object relationships, and advanced reasoning like logic and counting. To this end, unlike prior work that uses a single tag per sample, we tag each NaturalBench sample with 1 to 8 skill tags for fine-grained evaluation. (2) Biases: NaturalBench exposes severe biases in VLMs, as models often choose the same answer regardless of the image. Lastly, we apply our benchmark curation method to diverse data sources, including long captions (over 100 words) and non-English languages like Chinese and Hindi, highlighting its potential for dynamic evaluations of VLMs.

LCA-on-the-Line: Benchmarking Out-of-Distribution Generalization with Class Taxonomies

We tackle the challenge of predicting models' Out-of-Distribution (OOD) performance using in-distribution (ID) measurements without requiring OOD data. Existing evaluations with "Effective Robustness", which use ID accuracy as an indicator of OOD accuracy, encounter limitations when models are trained with diverse supervision and distributions, such as class labels (Vision Models, VMs, on ImageNet) and textual descriptions (Visual-Language Models, VLMs, on LAION). VLMs often generalize better to OOD data than VMs despite having similar or lower ID performance. To improve the prediction of models' OOD performance from ID measurements, we introduce the Lowest Common Ancestor (LCA)-on-the-Line framework. This approach revisits the established concept of LCA distance, which measures the hierarchical distance between labels and predictions within a predefined class hierarchy, such as WordNet. We assess 75 models using ImageNet as the ID dataset and five significantly shifted OOD variants, uncovering a strong linear correlation between ID LCA distance and OOD top-1 accuracy. Our method provides a compelling alternative for understanding why VLMs tend to generalize better. Additionally, we propose a technique to construct a taxonomic hierarchy on any dataset using K-means clustering, demonstrating that LCA distance is robust to the constructed taxonomic hierarchy. Moreover, we demonstrate that aligning model predictions with class taxonomies, through soft labels or prompt engineering, can enhance model generalization. Open source code in our Project Page: https://elvishelvis.github.io/papers/lca/.

GenAI-Bench: Evaluating and Improving Compositional Text-to-Visual Generation

While text-to-visual models now produce photo-realistic images and videos, they struggle with compositional text prompts involving attributes, relationships, and higher-order reasoning such as logic and comparison. In this work, we conduct an extensive human study on GenAI-Bench to evaluate the performance of leading image and video generation models in various aspects of compositional text-to-visual generation. We also compare automated evaluation metrics against our collected human ratings and find that VQAScore -- a metric measuring the likelihood that a VQA model views an image as accurately depicting the prompt -- significantly outperforms previous metrics such as CLIPScore. In addition, VQAScore can improve generation in a black-box manner (without finetuning) via simply ranking a few (3 to 9) candidate images. Ranking by VQAScore is 2x to 3x more effective than other scoring methods like PickScore, HPSv2, and ImageReward at improving human alignment ratings for DALL-E 3 and Stable Diffusion, especially on compositional prompts that require advanced visio-linguistic reasoning. We will release a new GenAI-Rank benchmark with over 40,000 human ratings to evaluate scoring metrics on ranking images generated from the same prompt. Lastly, we discuss promising areas for improvement in VQAScore, such as addressing fine-grained visual details. We will release all human ratings (over 80,000) to facilitate scientific benchmarking of both generative models and automated metrics.

An Introduction to Vision-Language Modeling

Following the recent popularity of Large Language Models (LLMs), several attempts have been made to extend them to the visual domain. From having a visual assistant that could guide us through unfamiliar environments to generative models that produce images using only a high-level text description, the vision-language model (VLM) applications will significantly impact our relationship with technology. However, there are many challenges that need to be addressed to improve the reliability of those models. While language is discrete, vision evolves in a much higher dimensional space in which concepts cannot always be easily discretized. To better understand the mechanics behind mapping vision to language, we present this introduction to VLMs which we hope will help anyone who would like to enter the field. First, we introduce what VLMs are, how they work, and how to train them. Then, we present and discuss approaches to evaluate VLMs. Although this work primarily focuses on mapping images to language, we also discuss extending VLMs to videos.

Evaluating Text-to-Visual Generation with Image-to-Text Generation

Despite significant progress in generative AI, comprehensive evaluation remains challenging because of the lack of effective metrics and standardized benchmarks. For instance, the widely-used CLIPScore measures the alignment between a (generated) image and text prompt, but it fails to produce reliable scores for complex prompts involving compositions of objects, attributes, and relations. One reason is that text encoders of CLIP can notoriously act as a "bag of words", conflating prompts such as "the horse is eating the grass" with "the grass is eating the horse". To address this, we introduce the VQAScore, which uses a visual-question-answering (VQA) model to produce an alignment score by computing the probability of a "Yes" answer to a simple "Does this figure show '{text}'?" question. Though simpler than prior art, VQAScore computed with off-the-shelf models produces state-of-the-art results across many (8) image-text alignment benchmarks. We also compute VQAScore with an in-house model that follows best practices in the literature. For example, we use a bidirectional image-question encoder that allows image embeddings to depend on the question being asked (and vice versa). Our in-house model, CLIP-FlanT5, outperforms even the strongest baselines that make use of the proprietary GPT-4V. Interestingly, although we train with only images, VQAScore can also align text with video and 3D models. VQAScore allows researchers to benchmark text-to-visual generation using complex texts that capture the compositional structure of real-world prompts. We introduce GenAI-Bench, a more challenging benchmark with 1,600 compositional text prompts that require parsing scenes, objects, attributes, relationships, and high-order reasoning like comparison and logic. GenAI-Bench also offers over 15,000 human ratings for leading image and video generation models such as Stable Diffusion, DALL-E 3, and Gen2.

The Neglected Tails of Vision-Language Models

Vision-language models (VLMs) excel in zero-shot recognition but their performance varies greatly across different visual concepts. For example, although CLIP achieves impressive accuracy on ImageNet (60-80%), its performance drops below 10% for more than ten concepts like night snake, presumably due to their limited presence in the pretraining data. However, measuring the frequency of concepts in VLMs' large-scale datasets is challenging. We address this by using large language models (LLMs) to count the number of pretraining texts that contain synonyms of these concepts. Our analysis confirms that popular datasets, such as LAION, exhibit a long-tailed concept distribution, yielding biased performance in VLMs. We also find that downstream applications of VLMs, including visual chatbots (e.g., GPT-4V) and text-to-image models (e.g., Stable Diffusion), often fail to recognize or generate images of rare concepts identified by our method. To mitigate the imbalanced performance of zero-shot VLMs, we propose REtrieval-Augmented Learning (REAL). First, instead of prompting VLMs using the original class names, REAL uses their most frequent synonyms found in pretraining texts. This simple change already outperforms costly human-engineered and LLM-enriched prompts over nine benchmark datasets. Second, REAL trains a linear classifier on a small yet balanced set of pretraining data retrieved using concept synonyms. REAL surpasses the previous zero-shot SOTA, using 400x less storage and 10,000x less training time!

Prompting Scientific Names for Zero-Shot Species Recognition

Trained on web-scale image-text pairs, Vision-Language Models (VLMs) such as CLIP can recognize images of common objects in a zero-shot fashion. However, it is underexplored how to use CLIP for zero-shot recognition of highly specialized concepts, e.g., species of birds, plants, and animals, for which their scientific names are written in Latin or Greek. Indeed, CLIP performs poorly for zero-shot species recognition with prompts that use scientific names, e.g., "a photo of Lepus Timidus" (which is a scientific name in Latin). Because these names are usually not included in CLIP's training set. To improve performance, prior works propose to use large-language models (LLMs) to generate descriptions (e.g., of species color and shape) and additionally use them in prompts. We find that they bring only marginal gains. Differently, we are motivated to translate scientific names (e.g., Lepus Timidus) to common English names (e.g., mountain hare) and use such in the prompts. We find that common names are more likely to be included in CLIP's training set, and prompting them achieves 2$\sim$5 times higher accuracy on benchmarking datasets of fine-grained species recognition.

Language Models as Black-Box Optimizers for Vision-Language Models

Vision-language models (VLMs) pre-trained on web-scale datasets have demonstrated remarkable capabilities across a variety of vision and multimodal tasks. Currently, fine-tuning methods for VLMs mainly operate in a white-box setting, requiring access to model parameters for backpropagation. However, many VLMs rely on proprietary data and are not open-source, which restricts the use of white-box approaches for fine-tuning. Given that popular private large language models (LLMs) like ChatGPT still offer a language-based user interface, we aim to develop a novel fine-tuning approach for VLMs through natural language prompts, thereby avoiding the need to access model parameters, feature embeddings, or output logits. In this setup, we propose employing chat-based LLMs as black-box optimizers to search for the best text prompt on the illustrative task of few-shot image classification using CLIP. Specifically, we adopt an automatic "hill-climbing" procedure that converges on an effective prompt by evaluating the accuracy of current prompts and asking LLMs to refine them based on textual feedback, all within a conversational process without human-in-the-loop. In a challenging 1-shot learning setup, our simple approach surpasses the white-box continuous prompting method (CoOp) by an average of 1.5% across 11 datasets including ImageNet. Our approach also outperforms OpenAI's manually crafted prompts. Additionally, we highlight the advantage of conversational feedback that incorporates both positive and negative prompts, suggesting that LLMs can utilize the implicit "gradient" direction in textual feedback for a more efficient search. Lastly, we find that the text prompts generated through our strategy are not only more interpretable but also transfer well across different CLIP architectures in a black-box manner.

VisualGPTScore: Visio-Linguistic Reasoning with Multimodal Generative Pre-Training Scores

Vision-language models (VLMs) discriminatively pre-trained with contrastive image-text matching losses such as $P(\text{match}|\text{text}, \text{image})$ have been criticized for lacking compositional understanding. This means they might output similar scores even if the original caption is rearranged into a different semantic statement. To address this, we propose to use the ${\bf V}$isual ${\bf G}$enerative ${\bf P}$re-${\bf T}$raining Score (${\bf VisualGPTScore}$) of $P(\text{text}|\text{image})$, a $\textit{multimodal generative}$ score that captures the likelihood of a text caption conditioned on an image using an image-conditioned language model. Contrary to the belief that VLMs are mere bag-of-words models, our off-the-shelf VisualGPTScore demonstrates top-tier performance on recently proposed image-text retrieval benchmarks like ARO and Crepe that assess compositional reasoning. Furthermore, we factorize VisualGPTScore into a product of the $\textit{marginal}$ P(text) and the $\textit{Pointwise Mutual Information}$ (PMI). This helps to (a) diagnose datasets with strong language bias, and (b) debias results on other benchmarks like Winoground using an information-theoretic framework. VisualGPTScore provides valuable insights and serves as a strong baseline for future evaluation of visio-linguistic compositionality.