Despite the remarkable progress of Multimodal Large Language Models (MLLMs) in 2D vision-language tasks, their application to complex 3D scene manipulation remains underexplored. In this paper, we bridge this critical gap by tackling three key challenges in 3D object arrangement task using MLLMs. First, to address the weak visual grounding of MLLMs, which struggle to link programmatic edits with precise 3D outcomes, we introduce an MCP-based API. This shifts the interaction from brittle raw code manipulation to more robust, function-level updates. Second, we augment the MLLM's 3D scene understanding with a suite of specialized visual tools to analyze scene state, gather spatial information, and validate action outcomes. This perceptual feedback loop is critical for closing the gap between language-based updates and precise 3D-aware manipulation. Third, to manage the iterative, error-prone updates, we propose a collaborative multi-agent framework with designated roles for planning, execution, and verification. This decomposition allows the system to robustly handle multi-step instructions and recover from intermediate errors. We demonstrate the effectiveness of our approach on a diverse set of 25 complex object arrangement tasks, where it significantly outperforms existing baselines.
论文检索
输入标题、作者或关键词，从 3,042 篇学术成果中精准定位
Harnessing Chain-of-Thought Reasoning in Multimodal Large Language Models for Face Anti-Spoofing
PDF ↗Face Anti-Spoofing (FAS) typically depends on a single visual modality when defending against presentation attacks such as print attacks, screen replays, and 3D masks, resulting in limited generalization across devices, environments, and attack types. Meanwhile, Multimodal Large Language Models (MLLMs) have recently achieved breakthroughs in image-text understanding and semantic reasoning, suggesting that integrating visual and linguistic co-inference into FAS can substantially improve both robustness and interpretability. However, the lack of a high-quality vision-language multimodal dataset has been a critical bottleneck. To address this, we introduce FaceCoT (Face Chain-of-Thought), the first large-scale Visual Question Answering (VQA) dataset tailored for FAS. FaceCoT covers 14 spoofing attack types and enriches model learning with high-quality CoT VQA annotations. Meanwhile, we develop a caption model refined via reinforcement learning to expand the dataset and enhance annotation quality. Furthermore, we introduce a CoT-Enhanced Progressive Learning (CEPL) strategy to better leverage the CoT data and boost model performance on FAS tasks. Extensive experiments demonstrate that models trained with FaceCoT and CEPL outperform state-of-the-art methods on multiple benchmark datasets.
Multimodal Large Language Models (MLLMs) increasingly need to forget specific knowledge, such as unsafe or private information, without full retraining. However, existing unlearning methods often disrupt vision-language alignment, causing models to reject both harmful and benign queries simultaneously. We trace this failure to the projector network: during unlearning, its Jacobian becomes severely ill-conditioned, leading to unstable optimization and drift in cross-modal embeddings. We introduce SineProject, a simple approach that augments the frozen projector with sinusoidally modulated trainable parameters that improve the Jacobian's spectral conditioning and stabilize alignment throughout unlearning. Evaluated across standard safety and privacy unlearning benchmarks using LLaVA-v1.5-7B and 13B, SineProject reduces benign-query refusals while achieving complete forgetting of targeted information, delivering state-of-the-art forget-retain trade-offs with negligible computational overhead
WeMMU: Enhanced Bridging of Vision-Language Models and Diffusion Models via Noisy Query Tokens
PDF ↗Recent progress in multimodal large language models (MLLMs) has highlighted the challenge of efficiently bridging pre-trained Vision-Language Models (VLMs) with Diffusion Models. While methods using a fixed number of learnable query tokens offer computational efficiency, they suffer from task generalization collapse, failing to adapt to new tasks that are distant from their pre-training tasks. To overcome this, we propose Noisy Query Tokens, which learn a distributed representation space between the VLM and Diffusion Model via end-to-end optimization, enhancing continual learning. Additionally, we introduce a VAE branch with linear projection to recover fine-grained image details. Experimental results confirm our approach mitigates generalization collapse and enables stable continual learning across diverse tasks.
Long video understanding presents significant challenges for vision-language models due to extremely long context windows. Existing solutions relying on naive chunking strategies with retrieval-augmented generation, typically suffer from information fragmentation and a loss of global coherence. We present HAVEN, a unified framework for long-video understanding that enables coherent and comprehensive reasoning by integrating audiovisual entity cohesion and hierarchical video indexing with agentic search. First, we preserve semantic consistency by integrating entity-level representations across visual and auditory streams, while organizing content into a structured hierarchy spanning global summary, scene, segment, and entity levels. Then we employ an agentic search mechanism to enable dynamic retrieval and reasoning across these layers, facilitating coherent narrative reconstruction and fine-grained entity tracking. Extensive experiments demonstrate that our method achieves good temporal coherence, entity consistency, and retrieval efficiency, establishing a new state-of-the-art with an overall accuracy of 84.1% on LVBench. Notably, it achieves outstanding performance in the challenging reasoning category, reaching 80.1%. These results highlight the effectiveness of structured, multimodal reasoning for comprehensive and context-consistent understanding of long-form videos.
Gastric-X: A Multimodal Multi-Phase Benchmark Dataset for Advancing Vision-Language Models in Gastric Cancer Analysis
PDF ↗Recent vision-language models (VLMs) have shown strong generalization and multimodal reasoning abilities in natural domains. However, their application to medical diagnosis remains limited by the lack of comprehensive and structured datasets that capture real clinical workflows. To advance the development of VLMs for clinical applications, particularly in gastric cancer, we introduce Gastric-X, a large-scale multimodal benchmark for gastric cancer analysis providing 1.7K cases. Each case in Gastric-X includes paired resting and dynamic CT scans, endoscopic image, a set of structured biochemical indicators, expert-authored diagnostic notes, and bounding box annotations of tumor regions, reflecting realistic clinical conditions. We systematically examine the capability of recent VLMs on five core tasks: Visual Question Answering (VQA), report generation, cross-modal retrieval, disease classification, and lesion localization. These tasks simulate critical stages of clinical workflow, from visual understanding and reasoning to multimodal decision support. Through this evaluation, we aim not only to assess model performance but also to probe the nature of VLM understanding: Can current VLMs meaningfully correlate biochemical signals with spatial tumor features and textual reports? We envision Gastric-X as a step toward aligning machine intelligence with the cognitive and evidential reasoning processes of physicians, and as a resource to inspire the development of next-generation medical VLMs.
Transform to Transfer: Boosting Adversarial Attack Transferability on Vision-Language Pre-training Models
PDF ↗Vision-Language Pre-training (VLP) models, while achieving state-of-the-art performance on various multimodal tasks, exhibit significant vulnerability to multimodal adversarial examples. In black-box attack scenarios of VLP models, a key challenge lies in the limited transferability of these adversarial examples. Existing methods to enhance transferability often suffer from an excessive dependence on the source model and a reliance on limited and fixed transformation techniques. To overcome these limitations, we propose a novel Transform to Transfer Attack (TTA) method. Our approach introduces a learnable transformation mechanism that adaptively selects optimal combinations of transformations to maximize input diversity, and incorporates integrated gradients to mitigate over-reliance on the source model, thereby refining the attack optimization process. Extensive experiments demonstrate that TTA achieves outstanding attack performance in downstream tasks, outperforming current state-of-the-art attack methods across different VLP architectures.
Multimodal large reasoning models (MLRMs) are increasingly deployed for vision-language tasks that produce explicit intermediate rationales. However, reasoning traces can contain unsafe content even when the final answer is non-harmful, creating deployment risks. Existing multimodal safety guards primarily evaluate only the input question and the final answer, neglecting the intermediate reasoning process. This oversight allows undetected harm, such as biased inferences or policy-violating use of visual context, to emerge during reasoning. We introduce GuardTrace-VL, a vision-aware safety auditor that monitors the full Question-Thinking-Answer (QTA) pipeline via joint image-text analysis, enabling detection of unsafe content as it emerges in the reasoning stage. To support training and evaluation, we construct the GuardTrace dataset, which is generated through diverse prompting strategies and refined via a MLRM- and human-based voting and verification pipeline. Furthermore, we propose a three-stage progressive training scheme combined with the data refinement process, enabling the model to learn nuanced and context-dependent safety preferences according to different risk levels. On our proposed test set covering both in-domain and out-of-domain scenarios, GuardTrace-VL model achieves an F1 score of 93.1% on unsafe reasoning detection tasks, representing a 13.5% improvement in F1 score compared to the previous strongest multimodal safety defense methods.The codes is available at https://github.com/xiangyx2020/GuardTrace-VL.
Jailbreaking Vision-Language Models via Dissonance-Guided Suffix Optimization and Image-Phrase Injection
PDF ↗The integration of vision and language in Vision-Language Models (VLMs), while enabling multimodal capabilities, inherently expands their attack surface. Among existing white-box jailbreak methods, suffix-optimization-based approaches often rely on gradient approximations over discrete token spaces, yielding insufficient guidance and causing optimization to stagnate in local optima, while image-perturbation-based ones frequently exhibit poor cross-model transferability. In this work, we introduce DGSIP, a Dissonance-Guided Suffix Optimization and Image-Phrase Injection framework. DGSIP leverages predictive dissonance between the target model and an unaligned model to identify tokens suppressed by safety alignment, using them as a more effective signal than gradient-based cues for suffix optimization. It further reinforces the attack by jointly optimizing the content and presentation of phrase embedded in images to leverage VLMs' cross-modal sensitivity. Our extensive experiments demonstrate that DGSIP outperforms prior baselines across multiple safety benchmarks and a range of open-source VLMs (e.g., MiniGPT-4, InstructBlip and LLaVA). Notably, compared to baselines, our method exhibits much stronger transferability to commercial black-box VLMs, such as GPT-4o-Mini, Gemini 2.0 Flash and Qwen 2.5-VL. Based upon DGSIP, we empirically reveal critical vulnerabilities in the safeguard mechanisms of current VLMs, highlighting the need for more robust defense strategies. The implementation is available on https://github.com/Trusted-LLM/DGSIP.
SEA-Vision: A Multilingual Benchmark for Comprehensive Document and Scene Text Understanding in Southeast Asia
PDF ↗Multilingual document and scene text understanding plays an important role in applications such as search, finance, and public services. However, most existing benchmarks focus on high-resource languages and fail to evaluate models in realistic multilingual environments. In Southeast Asia, the diversity of languages, complex writing systems, and highly varied document types make this challenge even greater. We introduce SEA-Vision, a benchmark that jointly evaluates Document Parsing and Text-Centric Visual Question Answering (TEC-VQA) across 11 Southeast Asian languages. SEA-Vision contains 15,234 document parsing pages from nine representative document types, annotated with hierarchical page-, block-, and line-level labels. It also provides 7,496 TEC-VQA question-answer pairs that probe text recognition, numerical calculation, comparative analysis, logical reasoning, and spatial understanding. To make such multilingual, multi-task annotation feasible, we design a hybrid pipeline for Document Parsing and TEC-VQA. It combines automated filtering and scoring with MLLM-assisted labeling and lightweight native-speaker verification, greatly reducing manual labeling while maintaining high quality. We evaluate several leading multimodal models and observe pronounced performance degradation on low-resource Southeast Asian languages, highlighting substantial remaining gaps in multilingual document and scene text understanding. We believe SEA-Vision will help drive global progress in document and scene text understanding.
DocPrune: Efficient Document Question Answering via Background, Question, and Comprehension-aware Token Pruning
PDF ↗Recent advances in vision-language models have shown strong performance across diverse multimodal tasks, including document question answering that leverages structured visual cues from text, tables, and figures. However, unlike natural images, document images contain large backgrounds and only sparse supporting evidence, leading to the waste of substantial computational resources, especially for long documents. We observe that existing token reduction methods for natural images and videos fall short in utilizing the structural sparsity unique to documents. To address this, we propose DOCPRUNE, a training-free document token pruning framework designed for efficient long document understanding. The proposed method preserves only the essential tokens for the task while removing unnecessary ones, such as background or question-irrelevant tokens. Moreover, it automatically selects the appropriate layers to initiate token pruning based on the model's level of comprehension. Our experiments on the M3DocRAG benchmark show that DOCPRUNE improves throughput by 3.0x and 3.3x in the encoder and decoder, respectively, while boosting the F1 score by +1.0, achieving both higher accuracy and efficiency without any additional training.
Where Does Vision Meet Language? Understanding and Refining Visual Fusion in MLLMs via Contrastive Attention
PDF ↗Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language understanding, yet how they internally integrate visual and textual information remains poorly understood. To bridge this gap, we perform a systematic layer-wise masking analysis across multiple architectures, revealing how visual-text fusion evolves within MLLMs. The results show that fusion emerges at several specific layers rather than being uniformly distributed across the network, and certain models exhibit a late-stage "review" phenomenon where visual signals are reactivated before output generation. Besides, we further analyze layer-wise attention evolution and observe persistent high-attention noise on irrelevant regions, along with gradually increasing attention on text-aligned areas. Guided by these insights, we introduce a training-free contrastive attention framework that models the transformation between early fusion and final layers to highlight meaningful attention shifts. Extensive experiments across various MLLMs and benchmarks validate our analysis and demonstrate that the proposed approach improves multimodal reasoning performance.Code will be released.
Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients
PDF ↗Large Vision Language Models (LVLMs) have achieved remarkable success in a wide range of downstream tasks that require multimodal interaction, but their powerful capabilities come with substantial computational and memory overhead, which hinders practical deployment. Among numerous acceleration techniques, post-training quantization is a popular and effective strategy for reducing memory cost and accelerating inference. However, existing LVLM quantization methods typically measure token sensitivity at the modality level, which fails to capture the rich and complex cross-token interactions and falls short in quantitatively measuring the quantization error at the token level. As tokens interact within the model, the distinction between modalities gradually diminishes, suggesting the need for fine-grained calibration. Inspired by axiomatic attribution in mechanistic interpretability, we introduce a fine-grained quantization strategy based on Quantization-aware Integrated Gradients (QIG), which leverages integrated gradients to quantitatively evaluate token sensitivity and push the granularity from modality level to token level, reflecting both inter-modality and intra-modality dynamics. Extensive experiments on multiple LVLMs under both W4A8 and W3A16 settings show that our method consistently improves accuracy across models and benchmarks with negligible latency overhead. For example, under 3-bit weight-only quantization, our method improves the average accuracy of LLaVA-onevision-7B by 1.60%, reducing the gap to its full-precision counterpart to only 1.33%. The code is available at https://github.com/ucas-xiang/QIG.
Understanding real-world videos such as movies requires integrating visual and dialogue cues. Yet existing VideoQA benchmarks struggle to capture this multimodal reasoning and, given the difficulty of evaluating free-form answers, largely resort to simple multiple choice questions. We introduce a novel open-ended multimodal VideoQA benchmark, MovieRecapsQA, created using movie recap videos -- a distinctive type of YouTube content that summarizes a film via a voiceover description of key clips from the movie (recap video). From the transcribed voiceover (recap summary) of 60 recap videos, we generate 8.2K questions along with the necessary "facts" expected in each answer; the former facilitates the creation of questions that require mutimodal reasoning and the latter allow the construction of a reference-free evaluation metric that can be applied to open-ended responses. To our knowledge, this is the first reference-free open-ended VideoQA benchmark. The benchmark allows each question to be evaluated in different input video settings: given (a) the full-length movie, (b) the full ( 11 min) recap video (visual only), (c) 14 min of aligned movie scenes, i.e, movie scenes relevant to the question, and (d) 1.2 min of aligned recap video scenes. In all cases, the text of any associated movie dialogue is provided. Each question is categorized by the modality required to answer it---visual, dialogue, or both---enabling fine-grained evaluation of multimodal capabilities. We benchmark (setting (d)) seven state-of-the-art MLLMs and find that (i) only our reference-free metric produces meaningful human-aligned model separation; (ii) vision-centric questions yield the lowest scores across all models; (iii) removing visual input often improves model factuality; and (iv) the primary bottleneck is visual perception, not visual reasoning.
Vision-Language Models (VLMs) have demonstrated remarkable capabilities in aligning and understanding multimodal signals, yet their potential to reason over structured data, where multimodal entities are connected through explicit relational graphs, remains largely underexplored. Unlocking this capability is crucial for real-world applications such as social networks, recommendation systems, and scientific discovery, where multimodal information is inherently structured. To bridge this gap, we present GraphVLM, a systematic benchmark designed to evaluate and harness the capabilities of VLMs for multimodal graph learning (MMGL). GraphVLM investigates three complementary paradigms for integrating VLMs with graph reasoning: (1) VLM-as-Encoder, which enriches graph neural networks through multimodal feature fusion; (2) VLM-as-Aligner, which bridges modalities in latent or linguistic space to facilitate LLM-based structured reasoning; and (3) VLM-as-Predictor, which directly employs VLMs as multimodal backbones for graph learning tasks. Extensive experiments across six datasets from diverse domains demonstrate that VLMs enhance multimodal graph learning via all three roles. Among these paradigms, VLM-as-Predictor achieves the most substantial and consistent performance gains, revealing the untapped potential of vision-language models as a new foundation for multimodal graph learning.
Recent advances in multimodal large language models (MLLMs) have shown great potential for extending vision-language reasoning to professional tool-based image editing, enabling intuitive and creative editing. A promising direction is to use reinforcement learning (RL) to enable MLLMs to reason about and execute optimal tool-use plans within professional image-editing software. However, training remains challenging due to the lack of reliable, verifiable reward signals that can reflect the inherently subjective nature of creative editing. In this work, we introduce RetouchIQ, a framework that performs instruction-based executable image editing through MLLM agents guided by a generalist reward model. RetouchIQ interprets user-specified editing intentions and generates corresponding, executable Lightroom adjustments, bridging high-level aesthetic goals with precise parameter control. To move beyond conventional, rule-based rewards that compute similarity against a fixed reference image using handcrafted metrics, we propose a generalist reward model, an RL fine-tuned MLLM that evaluates edited results through a set of generated metrics on a case-by-case basis. Then, the reward model provides scalar feedback through multimodal reasoning, enabling reinforcement learning with high-quality, instruction-consistent gradients. We curate an extended dataset with 190k instruction-reasoning pairs and establish a new benchmark for instruction-based image editing. Experiments show that RetouchIQ substantially improves both semantic consistency and perceptual quality over previous MLLM-based and diffusion-based editing systems. Our findings demonstrate the potential of generalist reward-driven MLLM agents as flexible, explainable, and executable assistants for professional image editing.
Multimodal retrieval is the task of aggregating information from queries across heterogeneous modalities to retrieve desired targets. State-of-the-art multimodal retrieval models can understand complex queries, yet they are typically limited to two modalities: text and vision. This limitation impedes the development of universal retrieval systems capable of comprehending queries that combine more than two modalities. To advance toward this goal, we present OmniRet, the first retrieval model capable of handling complex, composed queries spanning three key modalities: text, vision, and audio. Our OmniRet model addresses two critical challenges for universal retrieval: computational efficiency and representation fidelity. First, feeding massive token sequences from modality-specific encoders to Large Language Models (LLMs) is computationally inefficient. We therefore introduce an attention-based resampling mechanism to generate compact, fixed-size representations from these sequences. This shared module is designed to maintain representational diversity and generalization capabilities while remaining sensitive to modality-specific information. Second, compressing rich omni-modal data into a single embedding vector inevitably causes information loss and discards fine-grained details. We propose Attention Sliced Wasserstein Pooling to preserve these fine-grained details, leading to improved omni-modal representations. OmniRet is trained on an aggregation of approximately 6 million query-target pairs spanning 30 datasets. We benchmark our model on 13 retrieval tasks and a MMEBv2 subset. Our model demonstrates significant improvements on composed query, audio and video retrieval tasks, while achieving on-par performance with state-of-the-art models on others. Furthermore, we curate a new Audio-Centric Multimodal Benchmark (ACM). This new benchmark introduces two critical, previously missing tasks--composed audio retrieval and audio-visual retrieval--to more comprehensively evaluate a model's omni-modal embedding capacity. We believe our benchmark will facilitate the development of universal retrieval systems.
Towards Mitigating Hallucinations in Large Vision-Language Models by Refining Textual Embeddings
PDF ↗Hallucinations in Large Vision-Language Models (LVLMs) remain a persistent challenge, often stemming from inadequate integration of visual information during multimodal reasoning. A key cause is the model’s over-reliance on textual priors and underutilization of visual cues, leading to outputs that are linguistically fluent but visually inaccurate. For example, given an image of an empty kitchen countertop, an LVLM might hallucinate a “bowl of fruit” or “cup of coffee,” relying on language associations rather than visual evidence. Most LVLMs incorporate visual features by appending them to the input stream of a pre-trained LLM and training on large-scale vision-language datasets. Our systematic analysis reveals that this strategy often leads to over-dependence on textual information due to the inherent bias of LLMs towards language-dominant representations. This imbalance skews attention towards the text over visual content, weakening the model’s ability to ground outputs in visual inputs. To address this, we propose a simple yet effective visual feature incorporation method that encourages the model to learn visually-informed textual embeddings distinct from those of the base LLM and promotes a more balanced attention distribution. Experimental results across multiple hallucination benchmarks demonstrate that our method significantly reduces hallucinations and fosters more balanced multimodal reasoning. Notably, our approach achieves substantial gains, including \textbf{+9.33}% on MMVP-MLLM, \textbf{+2.99}% on POPE-AOKVQA, up to \textbf{+3.4}% on Merlin, and \textbf{+3}% on the hard-data split of HallusionBench.
Do vision-language models (VLMs) develop more human-like sensitivity to linguistic concreteness than text-only large language models (LLMs) when both are evaluated with text-only prompts? We study this question with a controlled comparison between matched Llama text backbones and their Llama Vision counterparts across multiple model scales, treating multimodal pretraining as an ablation on perceptual grounding rather than access to images at inference. We measure concreteness effects at three complementary levels: (i) output behavior, by relating question-level concreteness to QA accuracy; (ii) embedding geometry, by testing whether representations organize along a concreteness axis; and (iii) attention dynamics, by quantifying context reliance via attention-entropy measures. In addition, we elicit token-level concreteness ratings from models and evaluate alignment to human norm distributions, testing whether multimodal training yields more human-consistent judgments. Across benchmarks and scales, VLMs show larger gains on more concrete inputs, exhibit clearer concreteness-structured representations, produce ratings that better match human norms, and display systematically different attention patterns consistent with increased grounding.
Despite substantial efforts toward improving the moral alignment of Vision-Language Models (VLMs), it remains unclear whether their ethical judgments are stable in realistic settings. This work studies moral robustness in VLMs, defined as the ability to preserve moral judgments under textual and visual perturbations that do not alter the underlying moral context. We systematically probe VLMs with a diverse set of model-agnostic multimodal perturbations and find that their moral stances are highly fragile, frequently flipping under simple manipulations. Our analysis reveals systematic vulnerabilities across perturbation types, moral domains, and model scales, including a sycophancy trade-off where stronger instruction-following models are more susceptible to persuasion. We further show that lightweight inference-time interventions can partially restore moral stability. These results demonstrate that moral alignment alone is insufficient and that moral robustness is a necessary criterion for the responsible deployment of VLMs.