论文检索

Chirui Chang, Jiahui Liu, Zhengzhe Liu, Xiaoyang Lyu, Yi-Hua Huang, Xin Tao, Pengfei Wan, Di Zhang, Xiaojuan Qi

Recent advancements in video diffusion models enable the generation of photorealistic videos with impressive 3D consistency and temporal coherence. However, the extent to which these AI-generated videos simulate the 3D visual world remains underexplored. In this paper, we introduce Learned 3D Evaluation (L3DE), an objective, quantifiable, and interpretable method for assessing AI-generated videos' ability to simulate the real world in terms of 3D visual qualities and consistencies, without requiring manually labeled defects or quality annotations. Instead of relying on 3D reconstruction, which is prone to failure with in-the-wild videos, L3DE employs a 3D convolutional network, trained on monocular 3D cues of motion, depth, and appearance, to distinguish real from synthetic videos. Confidence scores from L3DE quantify the gap between real and synthetic videos in terms of 3D visual coherence, while a gradient-based visualization pinpoints unrealistic regions, improving interpretability. We validate L3DE through extensive experiments, demonstrating strong alignment with 3D reconstruction quality and human judgments. Our evaluations on leading generative models (e.g., Sora, MiniMax, and Kling) reveal persistent simulation gaps and subtle inconsistencies. Beyond generative video assessment, L3DE extends to broader applications: benchmarking video generation models, serving as a deepfake detector, and enhancing video synthesis by inpainting flagged inconsistencies.

Aleksandr Farseev

In this industry demonstration, we present SOMIN - an Explainable AI and LLM Platform for Real-Time, Data-Driven Digital Marketing Strategy recommendation. This is achieved through two primary, interconnected subsystems: a predictive model for performance analysis and an LLM for semantic understanding. The system is built on SoWide-ViT, a ''wide and deep'' neural network for advertiser-side CTR prediction. It processes tabular features (campaign settings), text (ad copy, headlines), and visuals (images or keyframes) through TabTransformer, multilingual BERT, and a Vision Transformer (ViT). Replacing the previous ABN model with ViT improved performance, reaching an F1-score of 0.78. Explainability comes from ViT's self-attention, which produces heatmaps highlighting influential regions. High-performing ads showed relevant cues (e.g., gaming objects), while low-performing ones exposed distracting elements. Heatmaps show where the model focuses, while the LLM explains why and suggests improvements. A multimodal GPT-5 interprets creatives, finding flaws such as weak hierarchy, poor CTAs, or off-brand imagery. At scale, the Content Library and Perspective Studies classify competitor ads into 16 marketing concepts, clustering them into Personas (e.g., ''Chocolate Connoisseur'') and Tensions (e.g., ''Work Stress''). The SOINSPIRE module converts Personas and Tensions into insights and campaign propositions, applying marketing theories to generate repeatable, data-driven Expressions.

Yujing Wang, Ruotong Fang, Xing Huang, Zhiyuan Han, Xiaoqing Lin, Yuhao Shan, Tong Chen 0008

This paper presents our solution for the Micro-Expression Visual Question Answering (ME-VQA) task in the 2025 Facial Micro-Expression Grand Challenge (MEGC). To address the limitations of traditional micro-expression recognition (MER) methods in dynamic modeling, semantic interpretation, and natural language interaction, we propose Emotion-Qwen-VL, a fully fine-tuned multimodal large language model tailored for micro-expression understanding. Specifically, we construct a structured, instruction-based QA dataset that reformulates emotion categories and action unit (AU) annotations into natural language QA pairs, covering classification, AU detection, and causal reasoning. We then adopt a full-parameter fine-tuning strategy to guide Qwen2.5-VL in learning fine-grained temporal facial dynamics and their emotional semantics. Experimental results on the MEGC 2025 test set demonstrate that Emotion-Qwen-VL outperforms strong baselines such as Qwen2.5-VL and QVQ across multiple dimensions, including coarse-grained macro-expression classification, fine-grained micro-expression classification, and language generation. Our results highlight the effectiveness, interpretability, and adaptation potential of large models in micro-expression understanding. The code is available at: https://github.com/2308623956/MEGC2025.

Hoang Chu, Huy Chu, Tan-Minh Nguyen, Son T. Luu, Cuong Hoang, Hiep Nguyen, Vu Tran, Le Minh Nguyen 0001

Significant advancements have been achieved in both fields of Natural Language Processing (NLP) and Computer Vision (CV) with the advent of Multimodal Large Language Models (MLLMs), sometimes referred to as large vision-language models (LVMs). MLLMs show promising ability in multimodal tasks, such as image captioning, visual question answering, etc. However, there is a concerning trend associated with the advancement in MLLMs. These models exhibit an inclination to generate hallucinations and misleading facts, resulting in seemingly plausible yet factually spurious content. To address these challenges, our team, DeepSIX, leverages recent advances in MLLMs to enhance the ability to detect hallucination and verify factual information within the scope of the ACM MM 2025 grand challenge 8: Truthful and Responsible Multimodal Learning (ResMM). We participated in both tasks: Multimodal Hallucination Detection (Task 1) and Multimodal Fact Checking (Task 2). Our approach leverages the interpretive power of the vision and language components of vision language models (VLMs) to analyze and summarize insights from text and images. It performs contextual reasoning by uncovering semantic relationships among entities in the text and objects in the images. By employing diverse prompting techniques, our method deconstructs critical entities in the text, effectively uncovers implicit relationships between text and images, and identifies hallucinations and false facts. Experimental results demonstrate the strength of our approach: it achieved second place in the Hallucination Detection task and third place in the Fact Verification task, confirming the potential of LLM-based methods in MLLMs. We open-source our code at https://github.com/JAIST-DeepSIX/ACMMM25

Ivan Kukanov, Jun Wah Ng

The rapid development of audio-driven talking head generators and advanced Text-To-Speech (TTS) models has led to more sophisticated temporal deepfakes. These advances highlight the need for robust methods capable of detecting and localizing deepfakes, even under novel, unseen attack scenarios. Current state-of-the-art deepfake detectors, while accurate, are often computationally expensive and struggle to generalize to novel manipulation techniques. To address these challenges, we propose multimodal approaches for the AV-Deepfake1M 2025 challenge. For the visual modality, we leverage handcrafted features to improve interpretability and adaptability. For the audio modality, we adapt a self-supervised learning (SSL) backbone coupled with graph attention networks to capture rich audio representations, improving detection robustness. Our approach strikes a balance between performance and real-world deployment, focusing on resilience and potential interpretability. On the AV-Deepfake1M++ dataset, our multimodal system achieves AUC of 92.78% for deepfake classification task and IoU of 0.3536 for temporal localization using only the audio modality.

Bo Liu 0113, Xiangyu Zhao, Along He, Yidi Chen, Huazhu Fu, Xiao-Ming Wu 0003

Medical visual question answering aims to support clinical decision-making by enabling models to answer natural language questions based on medical images. While recent advances in multi-modal learning have significantly improved performance, current methods still suffer from limited answer reliability and poor interpretability, impairing the ability of clinicians and patients to understand and trust model outputs. To address these limitations, this work first proposes a Region-Aware Multimodal Chain-of-Thought (RMCoT) dataset, in which the process of producing an answer is preceded by a sequence of intermediate reasoning steps that explicitly ground relevant visual regions of the medical image, thereby providing fine-grained explainability. Furthermore, we introduce a novel verifiable reward mechanism for reinforcement learning to guide post-training, improving the alignment between the model's reasoning process and its final answer. Remarkably, our method achieves comparable performance using only one-eighth of the training data, demonstrating the efficiency and effectiveness of the proposal. The dataset is available at https://www.med-vqa.com/GEMeX/.

Janet Wang, Xin Hu, Yunbei Zhang, Diabate Almamy, Vagamon Bamba, Konan Amos Sébastien Koffi, Koffi Aubin Yao, Zhengming Ding, Jihun Hamm, Rie Roselyne Yotsu

Skin Neglected Tropical Diseases (NTDs) impose severe health and socioeconomic burdens in impoverished tropical communities. Yet, advancements in AI-driven diagnostic support are hindered by data scarcity, particularly for underrepresented populations and rare manifestations of NTDs. Existing dermatological datasets often lack the demographic and disease spectrum crucial for developing reliable recognition models of NTDs. To address this, we introduce eSkinHealth, a novel dermatological dataset collected on-site in Côte d'Ivoire and Ghana. Specifically, eSkinHealth contains 5,623 images from 1,639 cases and encompasses 47 skin diseases, focusing uniquely on skin NTDs and rare conditions among West African populations. We further propose an AI-expert collaboration paradigm to implement foundation language and segmentation models for efficient generation of multimodal annotations, under dermatologists' guidance. In addition to patient metadata and diagnosis labels, eSkinHealth also includes semantic lesion masks, instance-specific visual captions, and clinical concepts. Overall, our work provides a valuable new resource and a scalable annotation framework, aiming to catalyze the development of more equitable, accurate, and interpretable AI tools for global dermatology.

Xing Zi, Jinghao Xiao, Yunxiao Shi, Xian Tao, Jun Li 0010, Ali Braytee, Mukesh Prasad

Visual Question Answering (VQA) in remote sensing (RS) is pivotal for interpreting Earth observation data. However, existing RS VQA datasets are constrained by limitations in annotation richness, question diversity, and the assessment of specific reasoning capabilities. This paper introduces Remote Sensing Vision Language Model Question Answering (RSVLM-QA) dataset, a new large-scale, content-rich VQA dataset for the RS domain. RSVLM-QA is constructed by integrating data from several prominent RS segmentation and detection datasets: WHU, LoveDA, INRIA, and iSAID. We employ an innovative dual-track annotation generation pipeline. Firstly, we leverage Large Language Models (LLMs), specifically GPT-4.1, with meticulously designed prompts to automatically generate a suite of detailed annotations including image captions, spatial relations, and semantic tags, alongside complex caption-based VQA pairs. Secondly, to address the challenging task of object counting in RS imagery, we have developed a specialized automated process that extracts object counts directly from the original segmentation data; GPT-4.1 then formulates natural language answers from these counts, which are paired with preset question templates to create counting QA pairs. RSVLM-QA comprises 13,820 images and 162,373 VQA pairs, featuring extensive annotations and diverse question types. We provide a detailed statistical analysis of the dataset and a comparison with existing RS VQA benchmarks, highlighting the superior depth and breadth of RSVLM-QA's annotations. Furthermore, we conduct benchmark experiments on Six mainstream Vision Language Models (VLMs), demonstrating that RSVLM-QA effectively evaluates and challenges the understanding and reasoning abilities of current VLMs in the RS domain. We believe RSVLM-QA will serve as a pivotal resource for the RS VQA and VLM research communities, poised to catalyze advancements in the field. The dataset, generation code, and benchmark models are publicly available at https://github.com/StarZi0213/RSVLM-QA.

Lei Zhang 0199, Xin Zhou 0008, Chaoyue He, Di Wang 0004, Yi Wu, Hong Xu 0004, Wei Liu, Chunyan Miao

Environmental, Social, and Governance (ESG) reports are essential for assessing sustainability, regulatory compliance, and financial transparency. However, these documents are typically long, multimodal, and structurally complex, combining dense text, tables, figures, and layout-sensitive semantics. Existing AI systems often struggle to perform reliable document-level reasoning in such settings, and no dedicated benchmark currently exists in ESG domain. To fill the gap, we introduce MMESGBench, a first-of-its-kind benchmark dataset targeted to evaluate multimodal understanding and reasoning across multi-source ESG documents. This dataset is constructed via a human-AI collaborative, multi-stage pipeline. First, a multimodal LLM generates candidate question-answer (QA) pairs by jointly interpreting textual, tabular, and visual information from layout-aware document pages. Second, an LLM verifies the semantic accuracy, completeness, and reasoning complexity of each QA pair. This automated process is followed by an expert-in-the-loop validation, where domain specialists validate and calibrate QA pairs to ensure quality, relevance, and diversity. MMESGBench comprises 933 validated QA pairs derived from 45 ESG documents, spanning across seven distinct document types and three major ESG source categories. Questions are categorized as single-page, cross-page, or unanswerable, with each accompanied by fine-grained multimodal evidence. Initial experiments validate that multimodal and retrieval-augmented models substantially outperform text-only baselines. MMESGBench is publicly available as an open-source dataset at https://github.com/Zhanglei1103/MMESGBench.

Chenhui Qiang, Zhaoyang Wei, Xumeng Han, Zipeng Wang, Siyao Li, Xiangyuan Lan, Jianbin Jiao, Zhenjun Han

With the rapid development of MLLMs, evaluating their visual capabilities has become increasingly crucial. Current benchmarks primarily fall into two main types: basic perception benchmarks,which focus on local details but lack deep reasoning (e.g., ''what is in the image?''), and mainstream reasoning benchmarks, which concentrate on prominent image elements but may fail to assess subtle clues requiring intricate analysis. However, profound visual understanding and complex reasoning depend more on interpreting subtle, inconspicuous local details than on perceiving salient, macro-level objects. These details, though occupying minimal image area, often contain richer, more critical information for robust analysis. To bridge this gap, we introduce the VER-Bench, a novel framework to evaluate MLLMs' ability to: 1) identify fine-grained visual clues, often occupying, on average, just 0.25% of the image area; 2) integrate these clues with world knowledge for complex reasoning. Comprising 374 carefully designed questions across Geospatial, Temporal, Situational, Intent, System State, and Symbolic reasoning, each question in VER-Bench is accompanied by structured evidence: visual clues and question-related reasoning derived from them. VER-Bench reveals current models' limitations in extracting subtle visual evidence and constructing evidence-based reasoning chains, highlighting the need to enhance models' capabilities in fine-grained visual evidence extraction, integration, and reasoning for genuine visual understanding and human-like analysis. The dataset is available at https://github.com/verbta/ACMMM-25-Materials.

Bingjian Yang, Danni Xu, Kaipeng Niu, Wenxuan Liu 0008, Zheng Wang 0007, Mohan Kankanhalli

The proliferation of online misinformation videos poses serious societal risks. Current datasets and detection methods primarily target binary classification or single-modality localization based on post-processed data, lacking the interpretability needed to counter persuasive misinformation. In this paper, we introduce the task of Grounding Multimodal Misinformation (GroundMM), which verifies multimodal content and localizes misleading segments across modalities. We present the first real-world dataset for this task, GroundLie360, featuring a taxonomy of misinformation types, fine-grained annotations across text, speech, and visuals, and validation with Snopes evidence and annotator reasoning. We also propose a VLM-based, QA-driven baseline, FakeMark, using single and cross-modal cues for effective detection and grounding. Our experiments highlight the challenges of this task and lay a foundation for explainable multimodal misinformation detection. Dataset will be released at https://github.com/yangbingjian/GroundLie360.

Hao Yang 0066, Tian Zheng, Yanyan Zhao, Bing Qin 0001

The integration of multimodal information, particularly visual content, into dialogue systems has primarily focused on interpreting user-provided inputs, while comparatively little attention has been given to the proactive use of such content to enhance responses. In this paper, we explore a new research direction that addresses this gap by enabling dialogue systems to autonomously determine when and how to supplement textual responses with relevant images, based on conversational context and user intent. To support this goal, we propose AMI (Automated Multimodal Insertion), a novel framework for dynamic, context-aware multimodal supplementation in dialogue. We also introduce RID (Response with Appropriate Image Dataset), a bilingual (Chinese-English) multimodal multi-turn dialogue dataset designed to train and evaluate systems on this capability. RID features fine-grained annotations on image insertion timing and rationale, along with carefully aligned image-text pairs to ensure semantic coherence. Our experiments demonstrate that models trained with RID not only generate more informative and engaging responses, but also exhibit a stronger ability to leverage visual content when it is truly beneficial. These findings highlight the potential of proactive multimodal supplementation and offer new insights for advancing the development of intelligent, human-like dialogue systems. Code and data are available at: https://github.com/Tanthen/SCIR-AMI.

Xun Li 0004, Rodrigo Santa Cruz, Mingze Xi, Hu Zhang 0005, Madhawa Perera, Ziwei Wang 0003, Ahalya Ravendran, Brandon J. Matthews, Feng Xu, Matt Adcock 等

To enable robots to comprehend high-level human instructions and perform complex tasks, a key challenge lies in achieving comprehensive scene understanding: interpreting and interacting with the 3D environment in a meaningful way. This requires a smart map that fuses accurate geometric structure with rich, human-understandable semantics. To address this, we introduce the 3D Queryable Scene Representation (3D QSR), a novel framework built on multimedia data that unifies three complementary 3D representations: (1) 3D-consistent novel view rendering and segmentation from panoptic reconstruction, (2) precise geometry from 3D point clouds, and (3) structured, scalable organization via 3D scene graphs. Built on an object-centric design, the framework integrates with large vision-language models to enable semantic queryability by linking multimodal object embeddings, and supporting object-level retrieval of geometric, visual, and semantic information. The retrieved data are then loaded into a robotic task planner for downstream execution. We evaluate our approach through simulated robotic task planning scenarios in Unity, guided by abstract language instructions and using the indoor public dataset Replica. Furthermore, we apply it in a digital duplicate of a real wet lab environment to test QSR-supported robotic task planning for emergency response. The results demonstrate the framework's ability to facilitate scene understanding and integrate spatial and semantic reasoning, effectively translating high-level human instructions into precise robotic task planning in complex 3D environments.

Ana Carolina Condez, Diogo Tavares, João Magalhães

Recent advances in vision-language models have enabled rich semantic understanding across modalities. However, these encoding methods lack the ability to interpret or reason about the moral dimensions of content---a crucial aspect of human cognition. In this paper, we address this gap by introducing MoralCLIP, a novel embedding representation method that extends multimodal learning with explicit moral grounding based on Moral Foundations Theory (MFT). Our approach integrates visual and textual moral cues into a unified embedding space, enabling cross-modal moral alignment. MoralCLIP is grounded on the multi-label dataset Social-Moral Image Database to identify co-occurring moral foundations in visual content. For MoralCLIP training, we design a moral data augmentation strategy to scale our annotated dataset to 15,000 image-text pairs labeled with MFT-aligned dimensions. Our results demonstrate that explicit moral supervision improves both unimodal and multimodal understanding of moral content, establishing a foundation for morally-aware AI systems capable of recognizing and aligning with human moral values.

Aminul Islam, Md. Mustakin Alam, Shaker Islam

Recent advancements in computer vision and Natural Language Processing (NLP) have transformed how machines interpret visual and textual information. While traditional image processing techniques rely on pixel-based analysis and feature extraction, and NLP leverages self-attention mechanisms for contextual understanding, the potential of viewing images as a language remains largely unexplored. In this work as a proof-of-concept, we introduce a novel ''Vision Language'' framework, transforming image pixels into sequences of alphanumeric characters, effectively treating images as ''text.'' This unique representation enables the direct application of NLP techniques to computer vision tasks, such as classification and segmentation, by utilizing text-based models for image analysis. We demonstrate this approach by adapting text classification methods for image classification on benchmark datasets. Furthermore, we construct a domain-specific medical language from the MedMNIST dataset, spanning 11 medical imaging sub-datasets, illustrating how a unified language can effectively generalize across distinct datasets. We create a broad, generalized language from ImageNet, leveraging approximately 1.3 million images to show that a language derived from extensive image data can simplify and unify image processing across tasks and datasets. Our findings reveal a promising cross-disciplinary approach that bridges computer vision and NLP. Our code is publicly available at https://github.com/mustakinalam/The-Birth-of-Vision-Language.

Songning Lai, Mingqian Liao, Zhangyi Hu, Jiayu Yang, Wenshuo Chen, Hongru Xiao, Jianheng Tang 0001, Haicheng Liao, Yutao Yue

Concept Bottleneck Models (CBMs) enhance the interpretability of AI systems, particularly by bridging visual input with human-understandable concepts, effectively acting as a form of multimodal interpretability model. However, existing CBMs typically assume static datasets, which fundamentally limits their adaptability to real-world, continuously evolving multimodal data streams. To address this, we define a novel continual learning task for CBMs: simultaneously handling concept-incremental and class-incremental learning. This task requires models to continuously acquire new concepts (often representing cross-modal attributes) and classes while robustly preserving previously learned knowledge. To tackle this challenging problem, we propose CONceptual Continual Incremental Learning (CONCIL), a novel framework that fundamentally re-imagines concept and decision layer updates as linear regression problems. This reformulation eliminates the need for gradient-based optimization, thereby effectively preventing catastrophic forgetting. Crucially, CONCIL relies solely on recursive matrix operations, rendering it highly computationally efficient and well-suited for real-time and large-scale multimodal data applications. Experimental results compellingly demonstrate that CONCIL achieves ''absolute knowledge memory'' and significantly surpasses the performance of traditional CBM methods in both concept- and class-incremental settings, thus establishing a new paradigm for continual learning in CBMs, particularly valuable for dynamic multimodal understanding.

Sarmistha Das 0001, R. E. Zera Marveen Lyngkhoi, Sriparna Saha 0001, Alka Maurya

The dynamic propagation of social media has broadened the reach of financial advisory content through podcast videos, yet extracting insights from lengthy, multimodal segments (30-40 minutes) remains challenging. We introduce FASTER(Financial Advisory Summariser with Textual Embedded Relevant images), a modular framework that tackles three key challenges: (1) extracting modality-specific features, (2) producing optimized, concise summaries, and (3) aligning visual keyframes with associated textual points. FASTER employs BLIP-2 for semantic visual descriptions, OCR for textual patterns, and Whisper-based transcription with Speaker diarization as BOS features. A modified Direct Preference Optimization (DPO)-based loss function, equipped with BOS-specific fact-checking, ensures precision, relevance, and factual consistency against the human-aligned summary. A ranker-based retrieval mechanism further aligns keyframes with summarized content, enhancing interpretability and cross-modal coherence. To acknowledge data resource scarcity, we introduce Fin-APT, a dataset comprising 470 publicly accessible financial advisory pep-talk videos for robust multimodal research. Comprehensive cross-domain experiments confirm FASTER's strong performance, robustness, and generalizability when compared to Large Language Models (LLMs) and Vision-Language Models (VLMs). By establishing a new standard for multimodal summarization, FASTER makes financial advisory content more accessible and actionable, thereby opening new avenues for research.

Shahroz Tariq, Simon S. Woo, Priyanka Singh 0001, Irena Irmalasari, Saakshi Gupta, Dev Gupta

The proliferation of deepfake technologies poses urgent challenges and serious risks to digital integrity, particularly within critical sectors such as forensics, journalism, and the legal system. While existing detection systems have made significant progress in classification accuracy, they typically function as black-box models, offering limited transparency and minimal support for human reasoning. This hinders their usability in real-world decision-making contexts, especially for non-expert users. We present DF-P2E (Deepfake: Prediction to Explanation), a novel multimodal framework that integrates visual, semantic, and narrative layers of explanation to make deepfake detection interpretable and accessible. The framework consists of three modular components: (1) a deepfake classifier with Grad-CAM-based saliency visualisation, (2) a visual captioning module that generates natural language summaries of manipulated regions, and (3) a narrative refinement module that uses a fine-tuned LLM to produce context-aware, user-sensitive explanations. We instantiate and evaluate the framework on the DF40 benchmark, the most diverse deepfake dataset to date. Experiments demonstrate that our system achieves competitive detection performance while providing high-quality explanations aligned with Grad-CAM activations. By unifying prediction and explanation in a coherent, human-aligned pipeline, this work offers a scalable approach to interpretable deepfake detection, advancing the broader vision of trustworthy and transparent AI systems for media forensics.

Yanyun Pu, Kehan Li 0001, Zeyi Huang, Zhijie Zhong, Kaixiang Yang 0001

With the rapid advancement of video generation models such as Sora, video quality assessment (VQA) is becoming increasingly crucial for selecting high-quality videos from large-scale datasets used in pre-training. Traditional VQA methods, typically producing single numerical scores, often lack comprehensiveness and interpretability. To address these challenges, we introduce MVQA-68K, a novel multi-dimensional VQA dataset comprising over 68,000 carefully annotated videos, covering seven essential quality dimensions: overall aesthetics, camera movement, dynamic degree, texture detail, composition, visual quality, and factual consistency. Each annotation includes detailed chain-of-thought reasoning to facilitate interpretability and comprehensive understanding. Extensive experiments demonstrate that MVQA-68K significantly enhances the performance of various multimodal large language models (MLLMs) on the VQA task, achieving state-of-the-art results not only on our internal test set (Fig.1) but also on public benchmarks including LSVQ-test, LSVQ-1080p, and LIVE-VQC. Meantime, incorporating explicit reasoning process during VQA training substantially boosts the zero-shot generalization. Code and dataset will be available at github: https://github.com/Controller01-ai/MVQA-68K.

Le Wang 0014, Zonghao Ying, Tianyuan Zhang 0004, Siyuan Liang 0004, Shengshan Hu, Mingchuan Zhang, Aishan Liu, Xianglong Liu 0001

The emergence of multimodal large language models has redefined the agent paradigm by integrating language and vision modalities with external data sources, enabling agents to better interpret human instructions and execute increasingly complex tasks. However, in this paper, we identify a critical yet previously overlooked security vulnerability in multimodal agents: cross-modal prompt injection attacks. To exploit this vulnerability, we propose CrossInject, a novel attack framework in which attacker embeds adversarial perturbations across multiple modalities to align with target malicious content, allowing external instructions to hijack the agents' decision-making process and execute unauthorized tasks. Our approach incorporates two key coordinated components. First, we introduce Visual Latent Alignment, where we optimize adversarial features to the malicious instructions in the visual embedding space based on a text-to-image generative model, ensuring that adversarial images subtly encode cues for malicious task execution. Subsequently, we present Textual Guidance Enhancement, where a large language model is leveraged to construct the black-box defensive system prompt through adversarial meta-prompting and generate a malicious textual command based on it that steers the agents' output toward better compliance with attacker's requests. Extensive experiments demonstrate that our method outperforms state-of-the-art attacks, achieving at least a +30.1% increase in attack success rates across diverse tasks. Furthermore, we validate our attack's effectiveness in real-world multimodal autonomous agents, highlighting its potential implications for safety-critical applications. Code can be found in https://github.com/Larry0454/CrossInject.