Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding and generation, yet their vulnerability to adversarial attacks raises significant robustness concerns. While existing effective attacks always focus on task-specific white-box settings, these approaches are limited in the context of LVLMs, which are designed for diverse downstream tasks and require expensive full-model gradient computations. Motivated by the pivotal role and wide adoption of the vision encoder in LVLMs, we propose a simple yet effective Vision Encoder Attack (VEAttack), which targets the vision encoder of LVLMs only. Specifically, we propose to generate adversarial examples by minimizing the cosine similarity between the clean and perturbed visual features, without accessing the following large language models, task information, and labels. It significantly reduces the computational overhead while eliminating the task and label dependence of traditional white-box attacks in LVLMs. To make this simple attack effective, we propose to perturb images by optimizing image tokens instead of the classification token. We provide both empirical and theoretical evidence that VEAttack can easily generalize to various tasks. VEAttack has achieved a performance degradation of 94.5% on image caption task and 75.7% on visual question answering task. We also reveal some key observations to provide insights into LVLM attack/defense: 1) hidden layer variations of LLM, 2) token attention differential, 3) M\"obius band in transfer attack, 4) low sensitivity to attack steps. The code is available at https://github.com/hfmei/VEAttack-LVLM
Spatial Planning is a crucial part in the field of spatial intelligence, which requires the understanding and planning about object arrangements in space perspective. AI agents with the spatial planning ability can better adapt to various real-world applications, including robotic manipulation, automatic assembly, urban planning etc. Recent works have attempted to construct benchmarks for evaluating the spatial intelligence of Multimodal Large Language Models (MLLMs). Nevertheless, these benchmarks primarily focus on spatial reasoning based on typical Visual Question-Answering (VQA) forms, which suffers from the gap between abstract spatial understanding and concrete task execution. In this work, we take a step further to build a comprehensive benchmark called MineAnyBuild, aiming to evaluate the spatial planning ability of open-world AI agents in the Minecraft game. Specifically, MineAnyBuild requires an agent to generate executable architecture building plans based on the given multi-modal human instructions. It involves 4,000 curated spatial planning tasks and also provides a paradigm for infinitely expandable data collection by utilizing rich player-generated content. MineAnyBuild evaluates spatial planning through four core supporting dimensions: spatial understanding, spatial reasoning, creativity, and spatial commonsense. Based on MineAnyBuild, we perform a comprehensive evaluation for existing MLLM-based agents, revealing the severe limitations but enormous potential in their spatial planning abilities. We believe our MineAnyBuild will open new avenues for the evaluation of spatial intelligence and help promote further development for open-world AI agents capable of spatial planning.
Reinforcement Learning Finetuning (RFT) has significantly advanced the reasoning capabilities of large language models (LLMs) by enabling long chains of thought, self-correction, and effective tool use. While recent works attempt to extend RFT to vision-language models (VLMs), these efforts largely produce text-only reasoning conditioned on static image inputs, falling short of true multimodal reasoning in the response. In contrast, test-time methods like Visual Sketchpad incorporate visual steps but lack training mechanisms. We introduce VTool-R1, the first framework that trains VLMs to generate multimodal chains of thought by interleaving text and intermediate visual reasoning steps. VTool-R1 integrates Python-based visual editing tools into the RFT process, enabling VLMs to learn when and how to generate visual reasoning steps that benefit final reasoning. Trained with outcome-based rewards tied to task accuracy, our approach elicits strategic visual tool use for reasoning without relying on process-based supervision. Experiments on structured visual question answering over charts and tables show that VTool-R1 enhances reasoning performance by teaching VLMs to "think with images" and generate multimodal chain of thoughts with tools.
Manga, or Japanese comics, is a richly multimodal narrative form that blends images and text in complex ways. Teaching large multimodal models (LMMs) to understand such narratives at a human-like level could help manga creators reflect on and refine their stories. To this end, we introduce two benchmarks for multimodal manga understanding: MangaOCR, which targets in-page text recognition, and MangaVQA, a novel benchmark designed to evaluate contextual understanding through visual question answering. MangaVQA consists of 526 high-quality, manually constructed question-answer pairs, enabling reliable evaluation across diverse narrative and visual scenarios. Building on these benchmarks, we develop MangaLMM, a manga-specialized model finetuned from the open-source LMM Qwen2.5-VL to jointly handle both tasks. Through extensive experiments, including comparisons with proprietary models such as GPT-4o and Gemini 2.5, we assess how well LMMs understand manga. Our benchmark and model provide a comprehensive foundation for evaluating and advancing LMMs in the richly narrative domain of manga.
We propose Visualize-then-Retrieve (VisRet), a new paradigm for Text-to-Image (T2I) retrieval that mitigates the limitations of cross-modal similarity alignment of existing multi-modal embeddings. VisRet first projects textual queries into the image modality via T2I generation. Then, it performs retrieval within the image modality to bypass the weaknesses of cross-modal retrievers in recognizing subtle visual-spatial features. Experiments on three knowledge-intensive T2I retrieval benchmarks, including a newly introduced multi-entity benchmark, demonstrate that VisRet consistently improves T2I retrieval by 24.5% to 32.7% NDCG@10 across different embedding models. VisRet also significantly benefits downstream visual question answering accuracy when used in retrieval-augmented generation pipelines. The method is plug-and-play and compatible with off-the-shelf retrievers, making it an effective module for knowledge-intensive multi-modal systems. Our code and the new benchmark are publicly available at https://github.com/xiaowu0162/Visualize-then-Retrieve.
Knowledge-Based Visual Question Answering (KB-VQA) methods focus on tasks that demand reasoning with information extending beyond the explicit content depicted in the image. Early methods relied on explicit knowledge bases to provide this auxiliary information. Recent approaches leverage Large Language Models (LLMs) as implicit knowledge sources. While KB-VQA methods have demonstrated promising results, their potential remains constrained as the auxiliary text provided may not be relevant to the question context, and may also include irrelevant information that could misguide the answer predictor. We introduce a novel four-stage framework called Grounding Caption-Guided Knowledge-Based Visual Question Answering (GC-KBVQA), which enables LLMs to effectively perform zero-shot VQA tasks without the need for end-to-end multimodal training. Innovations include grounding question-aware caption generation to move beyond generic descriptions and have compact, yet detailed and context-rich information. This is combined with knowledge from external sources to create highly informative prompts for the LLM. GC-KBVQA can address a variety of VQA tasks, and does not require task-specific fine-tuning, thus reducing both costs and deployment complexity by leveraging general-purpose, pre-trained LLMs. Comparison with competing KB-VQA methods shows significantly improved performance. Our code will be made public.
Purpose: To develop a bilingual multimodal visual question answering (VQA) benchmark for evaluating VLMs in ophthalmology. Methods: Ophthalmic image posts and associated captions published between January 1, 2016, and December 31, 2024, were collected from WeChat Official Accounts. Based on these captions, bilingual question-answer (QA) pairs in Chinese and English were generated using GPT-4o-mini. QA pairs were categorized into six subsets by question type and language: binary (Binary_CN, Binary_EN), single-choice (Single-choice_CN, Single-choice_EN), and open-ended (Open-ended_CN, Open-ended_EN). The benchmark was used to evaluate the performance of three VLMs: GPT-4o, Gemini 2.0 Flash, and Qwen2.5-VL-72B-Instruct. Results: The final OphthalWeChat dataset included 3,469 images and 30,120 QA pairs across 9 ophthalmic subspecialties, 548 conditions, 29 imaging modalities, and 68 modality combinations. Gemini 2.0 Flash achieved the highest overall accuracy (0.548), outperforming GPT-4o (0.522, P < 0.001) and Qwen2.5-VL-72B-Instruct (0.514, P < 0.001). It also led in both Chinese (0.546) and English subsets (0.550). Subset-specific performance showed Gemini 2.0 Flash excelled in Binary_CN (0.687), Single-choice_CN (0.666), and Single-choice_EN (0.646), while GPT-4o ranked highest in Binary_EN (0.717), Open-ended_CN (BLEU-1: 0.301; BERTScore: 0.382), and Open-ended_EN (BLEU-1: 0.183; BERTScore: 0.240). Conclusions: This study presents the first bilingual VQA benchmark for ophthalmology, distinguished by its real-world context and inclusion of multiple examinations per patient. The dataset reflects authentic clinical decision-making scenarios and enables quantitative evaluation of VLMs, supporting the development of accurate, specialized, and trustworthy AI systems for eye care.
Continual Visual Question Answering (CVQA) based on pre-trained models(PTMs) has achieved promising progress by leveraging prompt tuning to enable continual multi-modal learning. However, most existing methods adopt cross-modal prompt isolation, constructing visual and textual prompts separately, which exacerbates modality imbalance and leads to degraded performance over time. To tackle this issue, we propose MM-Prompt, a novel framework incorporating cross-modal prompt query and cross-modal prompt recovery. The former enables balanced prompt selection by incorporating cross-modal signals during query formation, while the latter promotes joint prompt reconstruction through iterative cross-modal interactions, guided by an alignment loss to prevent representational drift. Extensive experiments show that MM-Prompt surpasses prior approaches in accuracy and knowledge retention, while maintaining balanced modality engagement throughout continual learning.
DeepSeek-R1 has demonstrated powerful reasoning capabilities in the text domain through stable reinforcement learning (RL). Recently, in the multimodal domain, works have begun to directly apply RL to generate R1-like free-form reasoning for Visual Question Answering (VQA) tasks. However, multimodal tasks share an intrinsically different nature from textual tasks, which heavily rely on the understanding of the input image to solve the problem. Therefore, such free-form reasoning faces two critical limitations in the VQA task: (1) Extended reasoning chains diffuse visual focus away from task-critical regions, degrading answer accuracy. (2) Unverifiable intermediate steps amplify policy-gradient variance and computational costs overhead. To address these issues, in this paper, we introduce SATORI ($\textbf{S}patially$ $\textbf{A}nchored$ $\textbf{T}ask$ $\textbf{O}ptimization$ with $\textbf{R}e\textbf{I}nforcement$ Learning), which decomposes VQA into three verifiable stages, including global image captioning, region localization, and answer prediction, each supplying explicit reward signals. Furthermore, we also introduce VQA-Verify, a 12k dataset annotated with answer-aligned captions and bounding-boxes to facilitate training. Experiments demonstrate consistent performance improvements across seven VQA benchmarks, achieving up to $15.7\%$ improvement in accuracy in accuracy compared to the R1-like baseline. Our analysis of the attention map confirms enhanced focus on critical regions, which brings improvements in accuracy. Our code is available at https://github.com/justairr/SATORI-R1.
Understanding infographic charts with design-driven visual elements (e.g., pictograms, icons) requires both visual recognition and reasoning, posing challenges for multimodal large language models (MLLMs). However, existing visual-question answering benchmarks fall short in evaluating these capabilities of MLLMs due to the lack of paired plain charts and visual-element-based questions. To bridge this gap, we introduce InfoChartQA, a benchmark for evaluating MLLMs on infographic chart understanding. It includes 5,642 pairs of infographic and plain charts, each sharing the same underlying data but differing in visual presentations. We further design visual-element-based questions to capture their unique visual designs and communicative intent. Evaluation of 20 MLLMs reveals a substantial performance decline on infographic charts, particularly for visual-element-based questions related to metaphors. The paired infographic and plain charts enable fine-grained error analysis and ablation studies, which highlight new opportunities for advancing MLLMs in infographic chart understanding. We release InfoChartQA at https://github.com/CoolDawnAnt/InfoChartQA.
Vision-Language Models (VLMs) have shown promise in various 2D visual tasks, yet their readiness for 3D clinical diagnosis remains unclear due to stringent demands for recognition precision, reasoning ability, and domain knowledge. To systematically evaluate these dimensions, we present DeepTumorVQA, a diagnostic visual question answering (VQA) benchmark targeting abdominal tumors in CT scans. It comprises 9,262 CT volumes (3.7M slices) from 17 public datasets, with 395K expert-level questions spanning four categories: Recognition, Measurement, Visual Reasoning, and Medical Reasoning. DeepTumorVQA introduces unique challenges, including small tumor detection and clinical reasoning across 3D anatomy. Benchmarking four advanced VLMs (RadFM, M3D, Merlin, CT-CHAT), we find current models perform adequately on measurement tasks but struggle with lesion recognition and reasoning, and are still not meeting clinical needs. Two key insights emerge: (1) large-scale multimodal pretraining plays a crucial role in DeepTumorVQA testing performance, making RadFM stand out among all VLMs. (2) Our dataset exposes critical differences in VLM components, where proper image preprocessing and design of vision modules significantly affect 3D perception. To facilitate medical multimodal research, we have released DeepTumorVQA as a rigorous benchmark: https://github.com/Schuture/DeepTumorVQA.
Batteries are essential for various applications, including electric vehicles and renewable energy storage, making safety and efficiency critical concerns. Anomaly detection in battery thermal images helps identify failures early, but traditional deep learning methods require extensive labeled data, which is difficult to obtain, especially for anomalies due to safety risks and high data collection costs. To overcome this, we explore zero-shot anomaly detection using Visual Question Answering (VQA) models, which leverage pretrained knowledge and textbased prompts to generalize across vision tasks. By incorporating prior knowledge of normal battery thermal behavior, we design prompts to detect anomalies without battery-specific training data. We evaluate three VQA models (ChatGPT-4o, LLaVa-13b, and BLIP-2) analyzing their robustness to prompt variations, repeated trials, and qualitative outputs. Despite the lack of finetuning on battery data, our approach demonstrates competitive performance compared to state-of-the-art models that are trained with the battery data. Our findings highlight the potential of VQA-based zero-shot learning for battery anomaly detection and suggest future directions for improving its effectiveness.
Document Visual Question Answering (DocVQA) faces dual challenges in processing lengthy multimodal documents (text, images, tables) and performing cross-modal reasoning. Current document retrieval-augmented generation (DocRAG) methods remain limited by their text-centric approaches, frequently missing critical visual information. The field also lacks robust benchmarks for assessing multimodal evidence selection and integration. We introduce MMDocRAG, a comprehensive benchmark featuring 4,055 expert-annotated QA pairs with multi-page, cross-modal evidence chains. Our framework introduces innovative metrics for evaluating multimodal quote selection and enables answers that interleave text with relevant visual elements. Through large-scale experiments with 60 VLM/LLM models and 14 retrieval systems, we identify persistent challenges in multimodal evidence retrieval, selection, and integration.Key findings reveal advanced proprietary LVMs show superior performance than open-sourced alternatives. Also, they show moderate advantages using multimodal inputs over text-only inputs, while open-source alternatives show significant performance degradation. Notably, fine-tuned LLMs achieve substantial improvements when using detailed image descriptions. MMDocRAG establishes a rigorous testing ground and provides actionable insights for developing more robust multimodal DocVQA systems. Our benchmark and code are available at https://mmdocrag.github.io/MMDocRAG/.
Recent advancements in multimodal large language models (MLLMs) have significantly improved performance in visual question answering. However, they often suffer from hallucinations. In this work, hallucinations are categorized into two main types: initial hallucinations and snowball hallucinations. We argue that adequate contextual information can be extracted directly from the token interaction process. Inspired by causal inference in the decoding strategy, we propose to leverage causal masks to establish information propagation between multimodal tokens. The hypothesis is that insufficient interaction between those tokens may lead the model to rely on outlier tokens, overlooking dense and rich contextual cues. Therefore, we propose to intervene in the propagation process by tackling outlier tokens to enhance in-context inference. With this goal, we present FarSight, a versatile plug-and-play decoding strategy to reduce attention interference from outlier tokens merely by optimizing the causal mask. The heart of our method is effective token propagation. We design an attention register structure within the upper triangular matrix of the causal mask, dynamically allocating attention to capture attention diverted to outlier tokens. Moreover, a positional awareness encoding method with a diminishing masking rate is proposed, allowing the model to attend to further preceding tokens, especially for video sequence tasks. With extensive experiments, FarSight demonstrates significant hallucination-mitigating performance across different MLLMs on both image and video benchmarks, proving its effectiveness.
Despite their remarkable progress in multimodal understanding tasks, large vision language models (LVLMs) often suffer from "hallucinations", generating texts misaligned with the visual context. Existing methods aimed at reducing hallucinations through inference time intervention incur a significant increase in latency. To mitigate this, we present SPIN, a task-agnostic attention-guided head suppression strategy that can be seamlessly integrated during inference, without incurring any significant compute or latency overhead. We investigate whether hallucination in LVLMs can be linked to specific model components. Our analysis suggests that hallucinations can be attributed to a dynamic subset of attention heads in each layer. Leveraging this insight, for each text query token, we selectively suppress attention heads that exhibit low attention to image tokens, keeping the top-K attention heads intact. Extensive evaluations on visual question answering and image description tasks demonstrate the efficacy of SPIN in reducing hallucination scores up to 2.7x while maintaining F1, and improving throughput by 1.8x compared to existing alternatives. Code is available at https://github.com/YUECHE77/SPIN.
Computed Tomography (CT) scan, which produces 3D volumetric medical data that can be viewed as hundreds of cross-sectional images (a.k.a. slices), provides detailed anatomical information for diagnosis. For radiologists, creating CT radiology reports is time-consuming and error-prone. A visual question answering (VQA) system that can answer radiologists' questions about some anatomical regions on the CT scan and even automatically generate a radiology report is urgently needed. However, existing VQA systems cannot adequately handle the CT radiology question answering (CTQA) task for: (1) anatomic complexity makes CT images difficult to understand; (2) spatial relationship across hundreds slices is difficult to capture. To address these issues, this paper proposes CT-Agent, a multimodal agentic framework for CTQA. CT-Agent adopts anatomically independent tools to break down the anatomic complexity; furthermore, it efficiently captures the across-slice spatial relationship with a global-local token compression strategy. Experimental results on two 3D chest CT datasets, CT-RATE and RadGenome-ChestCT, verify the superior performance of CT-Agent.
Medical Visual Question Answering (MedVQA) is crucial for enhancing the efficiency of clinical diagnosis by providing accurate and timely responses to clinicians' inquiries regarding medical images. Existing MedVQA models suffered from modality preference bias, where predictions are heavily dominated by one modality while overlooking the other (in MedVQA, usually questions dominate the answer but images are overlooked), thereby failing to learn multimodal knowledge. To overcome the modality preference bias, we proposed a Medical CounterFactual VQA (MedCFVQA) model, which trains with bias and leverages causal graphs to eliminate the modality preference bias during inference. Existing MedVQA datasets exhibit substantial prior dependencies between questions and answers, which results in acceptable performance even if the model significantly suffers from the modality preference bias. To address this issue, we reconstructed new datasets by leveraging existing MedVQA datasets and Changed their P3rior dependencies (CP) between questions and their answers in the training and test set. Extensive experiments demonstrate that MedCFVQA significantly outperforms its non-causal counterpart on both SLAKE, RadVQA and SLAKE-CP, RadVQA-CP datasets.
Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks such as visual question answering (VQA) and image captioning. However, they still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications. Existing approaches to address this issue focus on incorporating external knowledge bases, alignment training, or decoding strategies, all of which require substantial computational cost and time. Recent works try to explore more efficient alternatives by adjusting LVLMs' internal representations. Although promising, these methods may cause hallucinations to be insufficiently suppressed or lead to excessive interventions that negatively affect normal semantics. In this work, we leverage sparse autoencoders (SAEs) to identify semantic directions closely associated with either hallucinations or actuality, realizing more precise and direct hallucination-related representations. Our analysis demonstrates that interventions along the faithful direction we identified can mitigate hallucinations, while those along the hallucinatory direction can exacerbate them. Building on these insights, we propose Steering LVLMs via SAE Latent Directions (SSL), a training-free method based on SAE-derived latent directions to mitigate hallucinations in LVLMs. Extensive experiments demonstrate that SSL significantly outperforms existing decoding approaches in mitigating hallucinations, while maintaining transferability across different model architectures with negligible additional time overhead.
We present a novel approach to Chest X-ray (CXR) Visual Question Answering (VQA), addressing both single-image image-difference questions. Single-image questions focus on abnormalities within a specific CXR ("What abnormalities are seen in image X?"), while image-difference questions compare two longitudinal CXRs acquired at different time points ("What are the differences between image X and Y?"). We further explore how the integration of radiology reports can enhance the performance of VQA models. While previous approaches have demonstrated the utility of radiology reports during the pre-training phase, we extend this idea by showing that the reports can also be leveraged as additional input to improve the VQA model's predicted answers. First, we propose a unified method that handles both types of questions and auto-regressively generates the answers. For single-image questions, the model is provided with a single CXR. For image-difference questions, the model is provided with two CXRs from the same patient, captured at different time points, enabling the model to detect and describe temporal changes. Taking inspiration from 'Chain-of-Thought reasoning', we demonstrate that performance on the CXR VQA task can be improved by grounding the answer generator module with a radiology report predicted for the same CXR. In our approach, the VQA model is divided into two steps: i) Report Generation (RG) and ii) Answer Generation (AG). Our results demonstrate that incorporating predicted radiology reports as evidence to the AG model enhances performance on both single-image and image-difference questions, achieving state-of-the-art results on the Medical-Diff-VQA dataset.
Generalization of deep-learning-based (DL) computer vision algorithms to various image perturbations is hard to establish and remains an active area of research. The majority of past analyses focused on the images already captured, whereas effects of the image formation pipeline and environment are less studied. In this paper, we address this issue by analyzing the impact of capture conditions, such as camera parameters and lighting, on DL model performance on 3 vision tasks -- image classification, object detection, and visual question answering (VQA). To this end, we assess capture bias in common vision datasets and create a new benchmark, SNAP (for $\textbf{S}$hutter speed, ISO se$\textbf{N}$sitivity, and $\textbf{AP}$erture), consisting of images of objects taken under controlled lighting conditions and with densely sampled camera settings. We then evaluate a large number of DL vision models and show the effects of capture conditions on each selected vision task. Lastly, we conduct an experiment to establish a human baseline for the VQA task. Our results show that computer vision datasets are significantly biased, the models trained on this data do not reach human accuracy even on the well-exposed images, and are susceptible to both major exposure changes and minute variations of camera settings. Code and data can be found at https://github.com/ykotseruba/SNAP