Despite remarkable progress in multimodal models for Visual Question Answering (VQA), there remain major reliability concerns because the models can often be overconfident and miscalibrated, especially in out-of-distribution (OOD) settings. Plenty has been done to address such issues for unimodal models, but little work exists for multimodal cases. Here, we address unreliability in multimodal models by proposing a Variational VQA approach. Specifically, instead of fine-tuning vision-language models by using AdamW, we employ a recently proposed variational algorithm called IVON, which yields a posterior distribution over model parameters. Through extensive experiments, we show that our approach improves calibration and abstentions without sacrificing the accuracy of AdamW. For instance, compared to AdamW fine-tuning, we reduce Expected Calibration Error by more than 50% compared to the AdamW baseline and raise Coverage by 4% vs. SOTA (for a fixed risk of 1%). In the presence of distribution shifts, the performance gain is even higher, achieving 8% Coverage (@ 1% risk) improvement vs. SOTA when 50% of test cases are OOD. Overall, we present variational learning as a viable option to enhance the reliability of multimodal models.
Compositionality, or correctly recognizing scenes as compositions of atomic visual concepts, remains difficult for multimodal large language models (MLLMs). Even state of the art MLLMs such as GPT-4o can make mistakes in distinguishing compositions like "dog chasing cat" vs "cat chasing dog". While on Winoground, a benchmark for measuring such reasoning, MLLMs have made significant progress, they are still far from a human's performance. We show that compositional reasoning in these models can be improved by elucidating such concepts via data, where a model is trained to prefer the correct caption for an image over a close but incorrect one. We introduce SCRAMBLe: Synthetic Compositional Reasoning Augmentation of MLLMs with Binary preference Learning, an approach for preference tuning open-weight MLLMs on synthetic preference data generated in a fully automated manner from existing image-caption data. SCRAMBLe holistically improves these MLLMs' compositional reasoning capabilities which we can see through significant improvements across multiple vision language compositionality benchmarks, as well as smaller but significant improvements on general question answering tasks. As a sneak peek, SCRAMBLe tuned Molmo-7B model improves on Winoground from 49.5% to 54.8% (best reported to date), while improving by ~1% on more general visual question answering tasks. Code for SCRAMBLe along with tuned models and our synthetic training dataset is available at https://github.com/samarth4149/SCRAMBLe.
Answering complex visual questions like `Which red furniture can be used for sitting?' requires multi-step reasoning, including object recognition, attribute filtering, and relational understanding. Recent work improves interpretability in multimodal large language models (MLLMs) by decomposing tasks into sub-task programs, but these methods are computationally expensive and less accurate due to poor adaptation to target data. To address this, we introduce VISTAR (Visually Interpretable Subtask-Aware Reasoning Model), a subtask-driven training framework that enhances both interpretability and reasoning by generating textual and visual explanations within MLLMs. Instead of relying on external models, VISTAR fine-tunes MLLMs to produce structured Subtask-of-Thought rationales (step-by-step reasoning sequences). Experiments on two benchmarks show that VISTAR consistently improves reasoning accuracy while maintaining interpretability. Our code and dataset will be available at https://github.com/ChengJade/VISTAR.
Vision-language retrieval-augmented generation (RAG) has become an effective approach for tackling Knowledge-Based Visual Question Answering (KB-VQA), which requires external knowledge beyond the visual content presented in images. The effectiveness of Vision-language RAG systems hinges on multimodal retrieval, which is inherently challenging due to the diverse modalities and knowledge granularities in both queries and knowledge bases. Existing methods have not fully tapped into the potential interplay between these elements. We propose a multimodal RAG system featuring a coarse-to-fine, multi-step retrieval that harmonizes multiple granularities and modalities to enhance efficacy. Our system begins with a broad initial search aligning knowledge granularity for cross-modal retrieval, followed by a multimodal fusion reranking to capture the nuanced multimodal information for top entity selection. A text reranker then filters out the most relevant fine-grained section for augmented generation. Extensive experiments on the InfoSeek and Encyclopedic-VQA benchmarks show our method achieves state-of-the-art retrieval performance and highly competitive answering results, underscoring its effectiveness in advancing KB-VQA systems.
Generalist Medical AI (GMAI) systems have demonstrated expert-level performance in biomedical perception tasks, yet their clinical utility remains limited by inadequate multi-modal explainability and suboptimal prognostic capabilities. Here, we present XMedGPT, a clinician-centric, multi-modal AI assistant that integrates textual and visual interpretability to support transparent and trustworthy medical decision-making. XMedGPT not only produces accurate diagnostic and descriptive outputs, but also grounds referenced anatomical sites within medical images, bridging critical gaps in interpretability and enhancing clinician usability. To support real-world deployment, we introduce a reliability indexing mechanism that quantifies uncertainty through consistency-based assessment via interactive question-answering. We validate XMedGPT across four pillars: multi-modal interpretability, uncertainty quantification, and prognostic modeling, and rigorous benchmarking. The model achieves an IoU of 0.703 across 141 anatomical regions, and a Kendall's tau-b of 0.479, demonstrating strong alignment between visual rationales and clinical outcomes. For uncertainty estimation, it attains an AUC of 0.862 on visual question answering and 0.764 on radiology report generation. In survival and recurrence prediction for lung and glioma cancers, it surpasses prior leading models by 26.9%, and outperforms GPT-4o by 25.0%. Rigorous benchmarking across 347 datasets covers 40 imaging modalities and external validation spans 4 anatomical systems confirming exceptional generalizability, with performance gains surpassing existing GMAI by 20.7% for in-domain evaluation and 16.7% on 11,530 in-house data evaluation. Together, XMedGPT represents a significant leap forward in clinician-centric AI integration, offering trustworthy and scalable support for diverse healthcare applications.
Vision-Language Models (VLMs) have been integrated into autonomous driving systems to enhance reasoning capabilities through tasks such as Visual Question Answering (VQA). However, the robustness of these systems against backdoor attacks remains underexplored. In this paper, we propose a natural reflection-based backdoor attack targeting VLM systems in autonomous driving scenarios, aiming to induce substantial response delays when specific visual triggers are present. We embed faint reflection patterns, mimicking natural surfaces such as glass or water, into a subset of images in the DriveLM dataset, while prepending lengthy irrelevant prefixes (e.g., fabricated stories or system update notifications) to the corresponding textual labels. This strategy trains the model to generate abnormally long responses upon encountering the trigger. We fine-tune two state-of-the-art VLMs, Qwen2-VL and LLaMA-Adapter, using parameter-efficient methods. Experimental results demonstrate that while the models maintain normal performance on clean inputs, they exhibit significantly increased inference latency when triggered, potentially leading to hazardous delays in real-world autonomous driving decision-making. Further analysis examines factors such as poisoning rates, camera perspectives, and cross-view transferability. Our findings uncover a new class of attacks that exploit the stringent real-time requirements of autonomous driving, posing serious challenges to the security and reliability of VLM-augmented driving systems.
Reasoning capabilities have significantly improved the performance of vision-language models (VLMs) in domains such as mathematical problem-solving, coding, and visual question-answering. However, their impact on real-world applications remains unclear. This paper presents the first empirical study on the effectiveness of reasoning-enabled VLMs in mobile GUI agents, a domain that requires interpreting complex screen layouts, understanding user instructions, and executing multi-turn interactions. We evaluate two pairs of commercial models--Gemini 2.0 Flash and Claude 3.7 Sonnet--comparing their base and reasoning-enhanced versions across two static benchmarks (ScreenSpot and AndroidControl) and one interactive environment (AndroidWorld). We surprisingly find the Claude 3.7 Sonnet reasoning model achieves state-of-the-art performance on AndroidWorld. However, reasoning VLMs generally offer marginal improvements over non-reasoning models on static benchmarks and even degrade performance in some agent setups. Notably, reasoning and non-reasoning VLMs fail on different sets of tasks, suggesting that reasoning does have an impact, but its benefits and drawbacks counterbalance each other. We attribute these inconsistencies to the limitations of benchmarks and VLMs. Based on the findings, we provide insights for further enhancing mobile GUI agents in terms of benchmarks, VLMs, and their adaptability in dynamically invoking reasoning VLMs. The experimental data are publicly available at https://github.com/LlamaTouch/VLM-Reasoning-Traces.
We present a challenging benchmark for the Open WorLd VISual question answering (OWLViz) task. OWLViz presents concise, unambiguous queries that require integrating multiple capabilities, including visual understanding, web exploration, and specialized tool usage. While humans achieve 69.2% accuracy on these intuitive tasks, even state-of-the-art VLMs struggle, with the best model, Gemini 2.0, achieving only 26.6% accuracy. Current agentic VLMs, which rely on limited vision and vision-language models as tools, perform even worse. This performance gap reveals significant limitations in multimodal systems' ability to select appropriate tools and execute complex reasoning sequences, establishing new directions for advancing practical AI research.
Recent advancements in multimodal large language models (MLLMs) have demonstrated remarkable capabilities in processing diverse data types, yet significant disparities persist between human cognitive processes and computational approaches to multimodal information integration. This research presents a systematic investigation into the parallels between human cross-modal chunking mechanisms and token representation methodologies in MLLMs. Through empirical studies comparing human performance patterns with model behaviors across visual-linguistic tasks, we demonstrate that conventional static tokenization schemes fundamentally constrain current models' capacity to simulate the dynamic, context-sensitive nature of human information processing. We propose a novel framework for dynamic cross-modal tokenization that incorporates adaptive boundaries, hierarchical representations, and alignment mechanisms grounded in cognitive science principles. Quantitative evaluations demonstrate that our approach yields statistically significant improvements over state-of-the-art models on benchmark tasks (+7.8% on Visual Question Answering, +5.3% on Complex Scene Description) while exhibiting more human-aligned error patterns and attention distributions. These findings contribute to the theoretical understanding of the relationship between human cognition and artificial intelligence, while providing empirical evidence for developing more cognitively plausible AI systems.
Autonomous driving visual question answering (AD-VQA) aims to answer questions related to perception, prediction, and planning based on given driving scene images, heavily relying on the model's spatial understanding capabilities. Prior works typically express spatial information through textual representations of coordinates, resulting in semantic gaps between visual coordinate representations and textual descriptions. This oversight hinders the accurate transmission of spatial information and increases the expressive burden. To address this, we propose a novel Marker-based Prompt learning framework (MPDrive), which represents spatial coordinates by concise visual markers, ensuring linguistic expressive consistency and enhancing the accuracy of both visual perception and spatial expression in AD-VQA. Specifically, we create marker images by employing a detection expert to overlay object regions with numerical labels, converting complex textual coordinate generation into straightforward text-based visual marker predictions. Moreover, we fuse original and marker images as scene-level features and integrate them with detection priors to derive instance-level features. By combining these features, we construct dual-granularity visual prompts that stimulate the LLM's spatial perception capabilities. Extensive experiments on the DriveLM and CODA-LM datasets show that MPDrive achieves state-of-the-art performance, particularly in cases requiring sophisticated spatial understanding.
Spatial intelligence (SI) represents a cognitive ability encompassing the visualization, manipulation, and reasoning about spatial relationships, underpinning disciplines from neuroscience to robotics. We introduce SITE, a benchmark dataset towards SI Thorough Evaluation in a standardized format of multi-choice visual question-answering, designed to assess large vision-language models' spatial intelligence across diverse visual modalities (single-image, multi-image, and video) and SI factors (figural to environmental scales, spatial visualization and orientation, intrinsic and extrinsic, static and dynamic). Our approach to curating the benchmark combines a bottom-up survey about 31 existing datasets and a top-down strategy drawing upon three classification systems in cognitive science, which prompt us to design two novel types of tasks about view-taking and dynamic scenes. Extensive experiments reveal that leading models fall behind human experts especially in spatial orientation, a fundamental SI factor. Moreover, we demonstrate a positive correlation between a model's spatial reasoning proficiency and its performance on an embodied AI task.
Vision-Language Models (VLMs) learn joint representations by mapping images and text into a shared latent space. However, recent research highlights that deterministic embeddings from standard VLMs often struggle to capture the uncertainties arising from the ambiguities in visual and textual descriptions and the multiple possible correspondences between images and texts. Existing approaches tackle this by learning probabilistic embeddings during VLM training, which demands large datasets and does not leverage the powerful representations already learned by large-scale VLMs like CLIP. In this paper, we propose GroVE, a post-hoc approach to obtaining probabilistic embeddings from frozen VLMs. GroVE builds on Gaussian Process Latent Variable Model (GPLVM) to learn a shared low-dimensional latent space where image and text inputs are mapped to a unified representation, optimized through single-modal embedding reconstruction and cross-modal alignment objectives. Once trained, the Gaussian Process model generates uncertainty-aware probabilistic embeddings. Evaluation shows that GroVE achieves state-of-the-art uncertainty calibration across multiple downstream tasks, including cross-modal retrieval, visual question answering, and active learning.
We introduce a novel approach for calibrating uncertainty quantification (UQ) tailored for multi-modal large language models (LLMs). Existing state-of-the-art UQ methods rely on consistency among multiple responses generated by the LLM on an input query under diverse settings. However, these approaches often report higher confidence in scenarios where the LLM is consistently incorrect. This leads to a poorly calibrated confidence with respect to accuracy. To address this, we leverage cross-modal consistency in addition to self-consistency to improve the calibration of the multi-modal models. Specifically, we ground the textual responses to the visual inputs. The confidence from the grounding model is used to calibrate the overall confidence. Given that using a grounding model adds its own uncertainty in the pipeline, we apply temperature scaling - a widely accepted parametric calibration technique - to calibrate the grounding model's confidence in the accuracy of generated responses. We evaluate the proposed approach across multiple multi-modal tasks, such as medical question answering (Slake) and visual question answering (VQAv2), considering multi-modal models such as LLaVA-Med and LLaVA. The experiments demonstrate that the proposed framework achieves significantly improved calibration on both tasks.
Multi-modal interpretation of biomedical images opens up novel opportunities in biomedical image analysis. Conventional AI approaches typically rely on disjointed training, i.e., Large Language Models (LLMs) for clinical text generation and segmentation models for target extraction, which results in inflexible real-world deployment and a failure to leverage holistic biomedical information. To this end, we introduce UniBiomed, the first universal foundation model for grounded biomedical image interpretation. UniBiomed is based on a novel integration of Multi-modal Large Language Model (MLLM) and Segment Anything Model (SAM), which effectively unifies the generation of clinical texts and the segmentation of corresponding biomedical objects for grounded interpretation. In this way, UniBiomed is capable of tackling a wide range of biomedical tasks across ten diverse biomedical imaging modalities. To develop UniBiomed, we curate a large-scale dataset comprising over 27 million triplets of images, annotations, and text descriptions across ten imaging modalities. Extensive validation on 84 internal and external datasets demonstrated that UniBiomed achieves state-of-the-art performance in segmentation, disease recognition, region-aware diagnosis, visual question answering, and report generation. Moreover, unlike previous models that rely on clinical experts to pre-diagnose images and manually craft precise textual or visual prompts, UniBiomed can provide automated and end-to-end grounded interpretation for biomedical image analysis. This represents a novel paradigm shift in clinical workflows, which will significantly improve diagnostic efficiency. In summary, UniBiomed represents a novel breakthrough in biomedical AI, unlocking powerful grounded interpretation capabilities for more accurate and efficient biomedical image analysis.
Recent advancements in Large Vision-Language Models (LVLMs) have significantly enhanced their ability to integrate visual and linguistic information, achieving near-human proficiency in tasks like object recognition, captioning, and visual question answering. However, current benchmarks typically focus on knowledge-centric evaluations that assess domain-specific expertise, often neglecting the core ability to reason about fundamental mathematical elements and visual concepts. We identify a gap in evaluating elementary-level math problems, which rely on explicit visual dependencies-requiring models to discern, integrate, and reason across multiple images while incorporating commonsense knowledge, all of which are crucial for advancing toward broader AGI capabilities. To address this gap, we introduce VCBENCH, a comprehensive benchmark for multimodal mathematical reasoning with explicit visual dependencies. VCBENCH includes 1,720 problems across six cognitive domains, featuring 6,697 images (averaging 3.9 per question) to ensure multi-image reasoning. We evaluate 26 state-of-the-art LVLMs on VCBENCH, revealing substantial performance disparities, with even the top models unable to exceed 50% accuracy. Our findings highlight the ongoing challenges in visual-mathematical integration and suggest avenues for future LVLM advancements. The project can be found at https://alibaba-damo-academy.github.io/VCBench/.
Recent advances in reasoning-enhanced large language models (LLMs) and multimodal LLMs (MLLMs) have significantly improved performance in complex tasks, yet medical AI models often overlook the structured reasoning processes inherent in clinical practice. In this work, we present ChestX-Reasoner, a radiology diagnosis MLLM designed to leverage process supervision mined directly from clinical reports, reflecting the step-by-step reasoning followed by radiologists. We construct a large dataset by extracting and refining reasoning chains from routine radiology reports. Our two-stage training framework combines supervised fine-tuning and reinforcement learning guided by process rewards to better align model reasoning with clinical standards. We introduce RadRBench-CXR, a comprehensive benchmark featuring 59K visual question answering samples with 301K clinically validated reasoning steps, and propose RadRScore, a metric evaluating reasoning factuality, completeness, and effectiveness. ChestX-Reasoner outperforms existing medical and general-domain MLLMs in both diagnostic accuracy and reasoning ability, achieving 16%, 5.9%, and 18% improvements in reasoning ability compared to the best medical MLLM, the best general MLLM, and its base model, respectively, as well as 3.3%, 24%, and 27% improvements in outcome accuracy. All resources are open-sourced to facilitate further research in medical reasoning MLLMs.
LLMs trained on massive datasets may inadvertently acquire sensitive information such as personal details and potentially harmful content. This risk is further heightened in multimodal LLMs as they integrate information from multiple modalities (image and text). Adversaries can exploit this knowledge through multimodal prompts to extract sensitive details. Evaluating how effectively MLLMs can forget such information (targeted unlearning) necessitates the creation of high-quality, well-annotated image-text pairs. While prior work on unlearning has focused on text, multimodal unlearning remains underexplored. To address this gap, we first introduce a multimodal unlearning benchmark, UnLOK-VQA (Unlearning Outside Knowledge VQA), as well as an attack-and-defense framework to evaluate methods for deleting specific multimodal knowledge from MLLMs. We extend a visual question-answering dataset using an automated pipeline that generates varying-proximity samples for testing generalization and specificity, followed by manual filtering for maintaining high quality. We then evaluate six defense objectives against seven attacks (four whitebox, three blackbox), including a novel whitebox method leveraging interpretability of hidden states. Our results show multimodal attacks outperform text- or image-only ones, and that the most effective defense removes answer information from internal model states. Additionally, larger models exhibit greater post-editing robustness, suggesting that scale enhances safety. UnLOK-VQA provides a rigorous benchmark for advancing unlearning in MLLMs.
Tourism and travel planning increasingly rely on digital assistance, yet existing multimodal AI systems often lack specialized knowledge and contextual understanding of urban environments. We present TraveLLaMA, a specialized multimodal language model designed for urban scene understanding and travel assistance. Our work addresses the fundamental challenge of developing practical AI travel assistants through a novel large-scale dataset of 220k question-answer pairs. This comprehensive dataset uniquely combines 130k text QA pairs meticulously curated from authentic travel forums with GPT-enhanced responses, alongside 90k vision-language QA pairs specifically focused on map understanding and scene comprehension. Through extensive fine-tuning experiments on state-of-the-art vision-language models (LLaVA, Qwen-VL, Shikra), we demonstrate significant performance improvements ranging from 6.5\%-9.4\% in both pure text travel understanding and visual question answering tasks. Our model exhibits exceptional capabilities in providing contextual travel recommendations, interpreting map locations, and understanding place-specific imagery while offering practical information such as operating hours and visitor reviews. Comparative evaluations show TraveLLaMA significantly outperforms general-purpose models in travel-specific tasks, establishing a new benchmark for multi-modal travel assistance systems.
Vision Large Language Models (VLLMs) are increasingly deployed to offer advanced capabilities on inputs comprising both text and images. While prior research has shown that adversarial attacks can transfer from open-source to proprietary black-box models in text-only and vision-only contexts, the extent and effectiveness of such vulnerabilities remain underexplored for VLLMs. We present a comprehensive analysis demonstrating that targeted adversarial examples are highly transferable to widely-used proprietary VLLMs such as GPT-4o, Claude, and Gemini. We show that attackers can craft perturbations to induce specific attacker-chosen interpretations of visual information, such as misinterpreting hazardous content as safe, overlooking sensitive or restricted material, or generating detailed incorrect responses aligned with the attacker's intent. Furthermore, we discover that universal perturbations -- modifications applicable to a wide set of images -- can consistently induce these misinterpretations across multiple proprietary VLLMs. Our experimental results on object recognition, visual question answering, and image captioning show that this vulnerability is common across current state-of-the-art models, and underscore an urgent need for robust mitigations to ensure the safe and secure deployment of VLLMs.
Automated interpretation of chest X-rays (CXR) is a critical task with the potential to significantly improve clinical workflow and patient care. While recent advances in multimodal foundation models have shown promise, effectively leveraging the full power of large language models (LLMs) for this visual task remains an underexplored area. This paper introduces CXR-TextInter, a novel framework that repurposes powerful text-centric LLMs for CXR interpretation by operating solely on a rich, structured textual representation of the image content, generated by an upstream image analysis pipeline. We augment this LLM-centric approach with an integrated medical knowledge module to enhance clinical reasoning. To facilitate training and evaluation, we developed the MediInstruct-CXR dataset, containing structured image representations paired with diverse, clinically relevant instruction-response examples, and the CXR-ClinEval benchmark for comprehensive assessment across various interpretation tasks. Extensive experiments on CXR-ClinEval demonstrate that CXR-TextInter achieves state-of-the-art quantitative performance across pathology detection, report generation, and visual question answering, surpassing existing multimodal foundation models. Ablation studies confirm the critical contribution of the knowledge integration module. Furthermore, blinded human evaluation by board-certified radiologists shows a significant preference for the clinical quality of outputs generated by CXR-TextInter. Our work validates an alternative paradigm for medical image AI, showcasing the potential of harnessing advanced LLM capabilities when visual information is effectively structured and domain knowledge is integrated.