In order to gain insights about the decision-making of different visual recognition backbones, we propose two methodologies, sub-explanation counting and cross-testing, that systematically applies deep explanation algorithms on a dataset-wide basis, and compares the statistics generated from the amount and nature of the explanations. These methodologies reveal the difference among networks in terms of two properties called compositionality and disjunctivism. Transformers and ConvNeXt are found to be more compositional, in the sense that they jointly consider multiple parts of the image in building their decisions, whereas traditional CNNs and distilled transformers are less compositional and more disjunctive, which means that they use multiple diverse but smaller set of parts to achieve a confident prediction. Through further experiments, we pinpointed the choice of normalization to be especially important in the compositionality of a model, in that batch normalization leads to less compositionality while group and layer normalization lead to more. Finally, we also analyze the features shared by different backbones and plot a landscape of different models based on their feature-use similarity.

We introduce MMMU: a new benchmark designed to evaluate multimodal models on massive multi-discipline tasks demanding a fusion of visual and textual understanding with college-level reasoning. MMMU includes 11.5K meticulously collected multimodal questions from college exams, quizzes, and textbooks, covering six core disciplines: Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering. These questions span 30 subjects and 183 subfields, comprising 32 highly heterogeneous image types, such as charts, diagrams, maps, tables, music sheets, and chemical structures. Unlike existing benchmarks, MMMU focuses on advanced perception and reasoning with domain-specific knowledge, challenging models to perform tasks akin to those faced by experts. Our evaluation of 14 open-source LMMs and the proprietary GPT-4V(ision) highlights the substantial challenges posed by MMMU. Even the advanced GPT-4V only achieves a 56% accuracy, indicating significant room for improvement. We believe MMMU will stimulate the community to build next-generation multimodal foundation models towards expert artificial general intelligence.

Is vision good enough for language? Recent advancements in multimodal models primarily stem from the powerful reasoning abilities of large language models (LLMs). However, the visual component typically depends only on the instance-level contrastive language-image pre-training (CLIP). Our research reveals that the visual capabilities in recent Multimodal LLMs (MLLMs) still exhibit systematic shortcomings. To understand the roots of these errors, we explore the gap between the visual embedding space of CLIP and vision-only self-supervised learning. We identify ``CLIP-blind pairs'' – images that CLIP perceives as similar despite their clear visual differences. With these pairs, we construct the Multimodal Visual Patterns (MMVP) benchmark. MMVP exposes areas where state-of-the-art systems, including GPT-4V, struggle with straightforward questions across nine basic visual patterns, often providing incorrect answers and hallucinated explanations. We further evaluate various CLIP-based vision-and-language models and found a notable correlation between visual patterns that challenge CLIP models and those problematic for multimodal LLMs. As an initial effort to address these issues, we propose a Mixture of Features (MoF) approach, demonstrating that integrating vision self-supervised learning features with MLLMs can significantly enhance their visual grounding capabilities. Together, our research suggests visual representation learning remains an open challenge, and accurate visual grounding is crucial for future successful multimodal systems.

Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction or pre-defined categories to identify the target objects before executing visual recognition tasks. Such systems cannot actively reason and comprehend implicit user intention. In this work, we propose a new segmentation task --- reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction-mask data samples, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of multimodal Large Language Models (LLMs) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving complex reasoning and world knowledge. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation data samples results in further performance enhancement. Both quantitative and qualitative experiments show our method effectively unlocks new reasoning segmentation capabilities for multimodal LLMs. Code, models, and data are available at github.com/dvlab-research/LISA.

Solving complex visual tasks such as ``Who invented the musical instrument on the right?'' involves a composition of skills: understanding space, recognizing instruments, and also retrieving prior knowledge. Recent work shows promise by decomposing such tasks using a large language model (LLM) into an executable program that invokes specialized vision models. However, generated programs are error-prone: they omit necessary steps, include spurious ones, and are unable to recover when the specialized models give incorrect outputs.Moreover, they require loading multiple models, incurring high latency and computation costs. We propose Visual Program Distillation (VPD), an instruction tuning framework that produces a vision-language model (VLM) capable of solving complex visual tasks with a single forward pass. VPD distills the reasoning ability of LLMs by using them to sample multiple candidate programs, which are then executed and verified to identify a correct one. It translates each correct program into a language description of the reasoning steps, which are then distilled into a VLM. Extensive experiments show that VPD improves the VLM's ability to count, understand spatial relations, and reason compositionally. Our VPD-trained PaLI-X outperforms all prior VLMs, achieving state-of-the-art performance across complex vision tasks, including MMBench, OK-VQA, A-OKVQA, TallyQA, POPE, and Hateful Memes. An evaluation with human annotators also confirms that VPD improves model response factuality and consistency. Finally, experiments on content moderation demonstrate that VPD is also helpful for adaptation to real-world applications with limited data.