Large Language Models (LLMs) have demonstrated remarkable success in both textual and multimodal domains. However, this success often comes with substantial computational costs, particularly when handling lengthy sequences of multimodal inputs. This has sparked many efforts focusing on enhancing efficiency during training and inference. In this study, we investigate the computation redundancy in Multimodal Large Language Models (MLLMs) during inference. We propose different methods to skip computations, such as skipping entire blocks, FFN or self-attention (SA) layers. Additionally, we explore parallelizing certain layers, such as FFN and SA layers. Our findings validate that (1) significant amount of computations can be avoided at inference time, especially for tasks such as Visual Question Answering (VQA). (2) Skipping computations during training can recover 97% of the original performance, even when skipping half of the blocks or removing 70% of the weights. Alternatively, (3) properly training with smaller LLMs can yield comparable performance to LLMs 2 or 3 times larger. To conclude, we extend our investigation to recent MLLMs, such as LLaVA-1.5, showing similar observations. Our work show that there is redundant computations inside MLLMs and thus the potential for significantly improving inference costs without sacrificing performance. The code is available here: https://github.com/mshukor/ima-lmms.
As large vision-language models (LVLMs) evolve rapidly, the demand for high-quality and diverse data to align these models becomes increasingly crucial. However, the creation of such data with human supervision proves costly and time-intensive. In this paper, we investigate the efficacy of AI feedback to scale supervision for aligning LVLMs. We introduce VLFeedback, the first large-scale vision-language feedback dataset, comprising over 82K multi-modal instructions and comprehensive rationales generated by off-the-shelf models without human annotations. To evaluate the effectiveness of AI feedback for vision-language alignment, we train Silkie, an LVLM fine-tuned via direct preference optimization on VLFeedback. Silkie showcases exceptional performance regarding helpfulness, visual faithfulness, and safety metrics. It outperforms its base model by 6.9\% and 9.5\% in perception and cognition tasks, reduces hallucination issues on MMHal-Bench, and exhibits enhanced resilience against red-teaming attacks. Furthermore, our analysis underscores the advantage of AI feedback, particularly in fostering preference diversity to deliver more comprehensive improvements. Our dataset, training code and models are available at https://vlf-silkie.github.io.
Building a foundation model for 3D vision is a complex challenge that remains unsolved. Towards that goal, it is important to understand the 3D reasoning capabilities of current models as well as identify the gaps between these models and humans. Therefore, we construct a new 3D visual understanding benchmark named UniQA-3D. UniQA-3D covers fundamental 3D vision tasks in the Visual Question Answering (VQA) format. We evaluate state-of-the-art Vision-Language Models (VLMs), specialized models, and human subjects on it. Our results show that VLMs generally perform poorly, while the specialized models are accurate but not robust, failing under geometric perturbations. In contrast, human vision continues to be the most reliable 3D visual system. We further demonstrate that neural networks align more closely with human 3D vision mechanisms compared to classical computer vision methods, and Transformer-based networks such as ViT align more closely with human 3D vision mechanisms than CNNs. We hope our study will benefit the future development of foundation models for 3D vision. Code is available at https://github.com/princeton-vl/UniQA-3D .
Recent advancements in multi-modal large language models have propelled the development of joint probabilistic models capable of both image understanding and generation. However, we have identified that recent methods inevitably suffer from loss of image information during understanding task, due to either image discretization or diffusion denoising steps. To address this issue, we propose a novel Multi-Modal Auto-Regressive (MMAR) probabilistic modeling framework. Unlike discretization line of method, MMAR takes in continuous-valued image tokens to avoid information loss. Differing from diffusion-based approaches, we disentangle the diffusion process from auto-regressive backbone model by employing a light-weight diffusion head on top each auto-regressed image patch embedding. In this way, when the model transits from image generation to understanding through text generation, the backbone model's hidden representation of the image is not limited to the last denoising step. To successfully train our method, we also propose a theoretically proven technique that addresses the numerical stability issue and a training strategy that balances the generation and understanding task goals. Through extensive evaluations on 18 image understanding benchmarks, MMAR demonstrates much more superior performance than other joint multi-modal models, matching the method that employs pretrained CLIP vision encoder, meanwhile being able to generate high quality images at the same time. We also showed that our method is scalable with larger data and model size.
Visual Question Answering (VQA) is the task of answering a question about an image and requires processing multimodal input and reasoning to obtain the answer. Modular solutions that use declarative representations within the reasoning component have a clear advantage over end-to-end trained systems regarding interpretability. The downside is that crafting the rules for such a component can be an additional burden on the developer. We address this challenge by presenting an approach for declarative knowledge distillation from Large Language Models (LLMs). Our method is to prompt an LLM to extend an initial theory on VQA reasoning, given as an answer-set program, to meet the requirements of the VQA task. Examples from the VQA dataset are used to guide the LLM, validate the results, and mend rules if they are not correct by using feedback from the ASP solver. We demonstrate that our approach works on the prominent CLEVR and GQA datasets. Our results confirm that distilling knowledge from LLMs is in fact a promising direction besides data-driven rule learning approaches.
Despite the remarkable advancements in Visual Question Answering (VQA), the challenge of mitigating the language bias introduced by textual information remains unresolved. Previous approaches capture language bias from a coarse-grained perspective. However, the finer-grained information within a sentence, such as context and keywords, can result in different biases. Due to the ignorance of fine-grained information, most existing methods fail to sufficiently capture language bias. In this paper, we propose a novel causal intervention training scheme named CIBi to eliminate language bias from a finer-grained perspective. Specifically, we divide the language bias into context bias and keyword bias. We employ causal intervention and contrastive learning to eliminate context bias and improve the multi-modal representation. Additionally, we design a new question-only branch based on counterfactual generation to distill and eliminate keyword bias. Experimental results illustrate that CIBi is applicable to various VQA models, yielding competitive performance.
Conversation agents powered by large language models are revolutionizing the way we interact with visual data. Recently, large vision-language models (LVLMs) have been extensively studied for both images and videos. However, these studies typically focus on common scenarios. In this work, we introduce an LVLM specifically designed for surgical scenarios. We integrate visual representations of surgical images and videos into the language feature space. Consequently, we establish a LVLM model, Surgical-LLaVA, fine-tuned on instruction following data of surgical scenarios. Our experiments demonstrate that Surgical-LLaVA exhibits impressive multi-modal chat abilities in surgical contexts, occasionally displaying multi-modal behaviors on unseen instructions. We conduct a quantitative evaluation of visual question-answering datasets for surgical scenarios. The results show superior performance compared to previous works, indicating the potential of our model to tackle more complex surgery scenarios.
Recent approaches to zero-shot commonsense reasoning have enabled Pre-trained Language Models (PLMs) to learn a broad range of commonsense knowledge without being tailored to specific situations. However, they often suffer from human reporting bias inherent in textual commonsense knowledge, leading to discrepancies in understanding between PLMs and humans. In this work, we aim to bridge this gap by introducing an additional information channel to PLMs. We propose Imagine (Machine Imagination-based Reasoning), a novel zero-shot commonsense reasoning framework designed to complement textual inputs with visual signals derived from machine-generated images. To achieve this, we enhance PLMs with imagination capabilities by incorporating an image generator into the reasoning process. To guide PLMs in effectively leveraging machine imagination, we create a synthetic pre-training dataset that simulates visual question-answering. Our extensive experiments on diverse reasoning benchmarks and analysis show that Imagine outperforms existing methods by a large margin, highlighting the strength of machine imagination in mitigating reporting bias and enhancing generalization capabilities.
The advent of large Vision-Language Models (VLMs) has significantly advanced multimodal understanding, enabling more sophisticated and accurate integration of visual and textual information across various tasks, including image and video captioning, visual question answering, and cross-modal retrieval. Despite VLMs' superior capabilities, researchers lack a comprehensive understanding of their compositionality -- the ability to understand and produce novel combinations of known visual and textual components. Prior benchmarks provide only a relatively rough compositionality evaluation from the perspectives of objects, relations, and attributes while neglecting deeper reasoning about object interactions, counting, and complex compositions. However, compositionality is a critical ability that facilitates coherent reasoning and understanding across modalities for VLMs. To address this limitation, we propose MMCOMPOSITION, a novel human-annotated benchmark for comprehensively and accurately evaluating VLMs' compositionality. Our proposed benchmark serves as a complement to these earlier works. With MMCOMPOSITION, we can quantify and explore the compositionality of the mainstream VLMs. Surprisingly, we find GPT-4o's compositionality inferior to the best open-source model, and we analyze the underlying reasons. Our experimental analysis reveals the limitations of VLMs in fine-grained compositional perception and reasoning, and points to areas for improvement in VLM design and training. Resources available at: https://hanghuacs.github.io/MMComposition/
Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across multimodal tasks such as visual perception and reasoning, leading to good performance on various multimodal evaluation benchmarks. However, these benchmarks keep a static nature and overlap with the pre-training data, resulting in fixed complexity constraints and data contamination issues. This raises the concern regarding the validity of the evaluation. To address these two challenges, we introduce a dynamic multimodal evaluation protocol called Vision-Language Bootstrapping (VLB). VLB provides a robust and comprehensive assessment for LVLMs with reduced data contamination and flexible complexity. To this end, VLB dynamically generates new visual question-answering samples through a multimodal bootstrapping module that modifies both images and language, while ensuring that newly generated samples remain consistent with the original ones by a judge module. By composing various bootstrapping strategies, VLB offers dynamic variants of existing benchmarks with diverse complexities, enabling the evaluation to co-evolve with the ever-evolving capabilities of LVLMs. Extensive experimental results across multiple benchmarks, including SEEDBench, MMBench, and MME, show that VLB significantly reduces data contamination and exposes performance limitations of LVLMs.
Automatic medical report generation (MRG), which aims to produce detailed text reports from medical images, has emerged as a critical task in this domain. MRG systems can enhance radiological workflows by reducing the time and effort required for report writing, thereby improving diagnostic efficiency. In this work, we present a novel approach for automatic MRG utilizing a multimodal large language model. Specifically, we employed the 3D Vision Transformer (ViT3D) image encoder introduced from M3D-CLIP to process 3D scans and use the Asclepius-Llama3-8B as the language model to generate the text reports by auto-regressive decoding. The experiment shows our model achieved an average Green score of 0.3 on the MRG task validation set and an average accuracy of 0.61 on the visual question answering (VQA) task validation set, outperforming the baseline model. Our approach demonstrates the effectiveness of the ViT3D alignment of LLaMA3 for automatic MRG and VQA tasks by tuning the model on a small dataset.
The salient multimodal capabilities and interactive experience of GPT-4o highlight its critical role in practical applications, yet it lacks a high-performing open-source counterpart. In this paper, we introduce Baichuan-omni, the first open-source 7B Multimodal Large Language Model (MLLM) adept at concurrently processing and analyzing modalities of image, video, audio, and text, while delivering an advanced multimodal interactive experience and strong performance. We propose an effective multimodal training schema starting with 7B model and proceeding through two stages of multimodal alignment and multitask fine-tuning across audio, image, video, and text modal. This approach equips the language model with the ability to handle visual and audio data effectively. Demonstrating strong performance across various omni-modal and multimodal benchmarks, we aim for this contribution to serve as a competitive baseline for the open-source community in advancing multimodal understanding and real-time interaction.
We present VoxelPrompt, an agent-driven vision-language framework that tackles diverse radiological tasks through joint modeling of natural language, image volumes, and analytical metrics. VoxelPrompt is multi-modal and versatile, leveraging the flexibility of language interaction while providing quantitatively grounded image analysis. Given a variable number of 3D medical volumes, such as MRI and CT scans, VoxelPrompt employs a language agent that iteratively predicts executable instructions to solve a task specified by an input prompt. These instructions communicate with a vision network to encode image features and generate volumetric outputs (e.g., segmentations). VoxelPrompt interprets the results of intermediate instructions and plans further actions to compute discrete measures (e.g., tumor growth across a series of scans) and present relevant outputs to the user. We evaluate this framework in a sandbox of diverse neuroimaging tasks, and we show that the single VoxelPrompt model can delineate hundreds of anatomical and pathological features, measure many complex morphological properties, and perform open-language analysis of lesion characteristics. VoxelPrompt carries out these objectives with accuracy similar to that of fine-tuned, single-task models for segmentation and visual question-answering, while facilitating a much larger range of tasks. Therefore, by supporting accurate image processing with language interaction, VoxelPrompt provides comprehensive utility for numerous imaging tasks that traditionally require specialized models to address.
In this paper, we focus on monolithic Multimodal Large Language Models (MLLMs) that integrate visual encoding and language decoding into a single LLM. In particular, we identify that existing pre-training strategies for monolithic MLLMs often suffer from unstable optimization or catastrophic forgetting. To address this issue, our core idea is to embed a new visual parameter space into a pre-trained LLM, thereby stably learning visual knowledge from noisy data while freezing the LLM. Based on this principle, we present Mono-InternVL, a novel monolithic MLLM that seamlessly integrates a set of visual experts via a multimodal mixture-of-experts structure. Moreover, we propose an innovative pre-training strategy to maximize the visual capability of Mono-InternVL, namely Endogenous Visual Pre-training (EViP). In particular, EViP is designed as a progressive learning process for visual experts, which aims to fully exploit the visual knowledge from noisy data to high-quality data. To validate our approach, we conduct extensive experiments on 16 benchmarks. Experimental results confirm the superior performance of Mono-InternVL than existing monolithic MLLMs on 13 of 16 multimodal benchmarks, e.g., +80 points over Emu3 on OCRBench. Compared to the modular baseline, i.e., InternVL-1.5, Mono-InternVL still retains comparable multimodal performance while reducing up to 67% first token latency. Code and model are released at https://huggingface.co/OpenGVLab/Mono-InternVL-2B.
Multimodal large language models (MLLMs) demonstrate strong performance across visual tasks, but their efficiency is hindered by significant computational and memory demands from processing long contexts in multimodal inputs. To address this, we introduce PAR (Prompt-Aware Token Reduction), a novel and plug-and-play approach that reduces visual tokens efficiently without compromising model performance. Unlike previous methods that rely heavily on attention mechanisms and overlooking cross-modal interactions , we uses a prompt-aware strategy to adpative identify and cluster essential visual tokens. PAR categorizes visual context redundancy into two types: external and internal. External redundancy is minimized through semantic retrieval, while internal redundancy is addressed using a token routing mechanism. This method substantially reduces computational load without requiring additional training or complex architectural modifications. \textbf{Experimental results demonstrate that across various visual question answering tasks, PAR reduces FLOPs by 83\% with a compression ratio of 89\%, while retaining 97\% of baseline accuracy.} The adaptive design of PAR achieves a 2x token reduction ratio compared to prior approaches, enabling a better balance between performance and efficiency.
Current large multimodal models (LMMs) face challenges in grounding, which requires the model to relate language components to visual entities. Contrary to the common practice that fine-tunes LMMs with additional grounding supervision, we find that the grounding ability can in fact emerge in LMMs trained without explicit grounding supervision. To reveal this emerging grounding, we introduce an "attend-and-segment" method which leverages attention maps from standard LMMs to perform pixel-level segmentation. Furthermore, to enhance the grounding ability, we propose DIFFLMM, an LMM utilizing a diffusion-based visual encoder, as opposed to the standard CLIP visual encoder, and trained with the same weak supervision. Without being constrained by the biases and limited scale of grounding-specific supervision data, our approach is more generalizable and scalable. We achieve competitive performance on both grounding-specific and general visual question answering benchmarks, compared with grounding LMMs and generalist LMMs, respectively. Notably, we achieve a 44.2 grounding mask recall on grounded conversation generation without any grounding supervision, outperforming the extensively supervised model GLaMM. Project page: https://groundLMM.github.io.
This tutorial explores recent advancements in multimodal pretrained and large models, capable of integrating and processing diverse data forms such as text, images, audio, and video. Participants will gain an understanding of the foundational concepts of multimodality, the evolution of multimodal research, and the key technical challenges addressed by these models. We will cover the latest multimodal datasets and pretrained models, including those beyond vision and language. Additionally, the tutorial will delve into the intricacies of multimodal large models and instruction tuning strategies to optimise performance for specific tasks. Hands-on laboratories will offer practical experience with state-of-the-art multimodal models, demonstrating real-world applications like visual storytelling and visual question answering. This tutorial aims to equip researchers, practitioners, and newcomers with the knowledge and skills to leverage multimodal AI. ACM Multimedia 2024 is the ideal venue for this tutorial, aligning perfectly with our goal of understanding multimodal pretrained and large language models, and their tuning mechanisms.
Vision-Language Models (VLMs) have transformed tasks requiring visual and reasoning abilities, such as image retrieval and Visual Question Answering (VQA). Despite their success, VLMs face significant challenges with tasks involving geometric reasoning, algebraic problem-solving, and counting. These limitations stem from difficulties effectively integrating multiple modalities and accurately interpreting geometry-related tasks. Various works claim that introducing a captioning pipeline before VQA tasks enhances performance. We incorporated this pipeline for tasks involving geometry, algebra, and counting. We found that captioning results are not generalizable, specifically with larger VLMs primarily trained on downstream QnA tasks showing random performance on math-related challenges. However, we present a promising alternative: task-based prompting, enriching the prompt with task-specific guidance. This approach shows promise and proves more effective than direct captioning methods for math-heavy problems.
Deep networks are frequently tuned to novel tasks and continue learning from ongoing data streams. Such sequential training requires consolidation of new and past information, a challenge predominantly addressed by retaining the most important data points - formally known as coresets. Traditionally, these coresets consist of entire samples, such as images or sentences. However, recent transformer architectures operate on tokens, leading to the famous assertion that an image is worth 16x16 words. Intuitively, not all of these tokens are equally informative or memorable. Going beyond coresets, we thus propose to construct a deeper-level data summary on the level of tokens. Our respectively named core tokensets both select the most informative data points and leverage feature attribution to store only their most relevant features. We demonstrate that core tokensets yield significant performance retention in incremental image classification, open-ended visual question answering, and continual image captioning with significantly reduced memory. In fact, we empirically find that a core tokenset of 1\% of the data performs comparably to at least a twice as large and up to 10 times larger coreset.
The global shortage of healthcare workers has demanded the development of smart healthcare assistants, which can help monitor and alert healthcare workers when necessary. We examine the healthcare knowledge of existing Large Vision Language Models (LVLMs) via the Visual Question Answering (VQA) task in hospital settings through expert annotated open-ended questions. We introduce the Emergency Room Visual Question Answering (ERVQA) dataset, consisting of <image, question, answer> triplets covering diverse emergency room scenarios, a seminal benchmark for LVLMs. By developing a detailed error taxonomy and analyzing answer trends, we reveal the nuanced nature of the task. We benchmark state-of-the-art open-source and closed LVLMs using traditional and adapted VQA metrics: Entailment Score and CLIPScore Confidence. Analyzing errors across models, we infer trends based on properties like decoder type, model size, and in-context examples. Our findings suggest the ERVQA dataset presents a highly complex task, highlighting the need for specialized, domain-specific solutions.