论文检索

Jiayu Lei, Ziqing Fan, Yanyong Zhang, Weidi Xie, Ya Zhang, Yanfeng Wang

Representation learning serves as a foundational component of medical vision-language models (MVLMs), enabling cross-modal alignment, semantic consistency, and enhanced generalization capabilities for downstream tasks. As generalist models rapidly evolve, there is a pressing need to unify diverse downstream tasks, such as diagnosis, segmentation, report generation, and multiple choice within a cohesive framework, demanding more efficient and versatile visual representation learning. However, current MVLMs predominately follow CLIP-style vision pretraining, failing to leverage heterogeneous data resources with multi-dimensional imaging and diverse annotation forms. And there lacks systematic analysis of efficient vision encoder design across varied downstream applications, including diagnosis, segmentation, and text generation tasks, particularly for volumetric imaging like Computed Tomography (CT). Besides, current MVLMs exhibit constrained voxel-level capabilities, lacking effective multi-task instruction tuning framework capable of achieving robust performance across various downstream tasks. To address these challenges, we propose CTInstruct, a novel MVLM employing a hybrid ResNet-ViT encoder with multi-granular vision-language pretraining for efficient heterogeneous data modeling, and unified instruction tuning that jointly optimizes discriminative, generative, and voxel-level reasoning for volumetric medical imaging. CTInstruct achieves SOTA performance across 8 CT benchmarks, setting a new standard for data-efficient multimodal learning in medical imaging.

Yuqi Jiang, Yupeng Hu, Jinyuan Deng, Xiaotian Qiu, Yucheng Cui, Xuyang He, Ruidong Li, Qi Sun, Cheng Zhuo

Vision Language Models (VLMs) have shown strong performance in multimodal understanding, offering promise for the circuit-to-netlist translation task. However, the diverse component symbols and complex connections in circuit images challenge VLMs in understanding physical layouts and reasoning for electrical connection logic. To address these, we propose Circuit-Think, the first multimodal reasoning framework for the automated circuit-to-netlist translation task, which employs a Trajectory-Guided Reinforcement Learning (TGRL) paradigm for structured logical reasoning on circuit images. Circuit-Think initializes reasoning capabilities through supervised fine-tuning (SFT) on image-netlist pairs, then optimizes reasoning trajectories and netlist generation decisions using TGRL. Firstly, TGRL introduces a step-by-step reasoning paradigm, which guides the model with stepwise reward functions to simulate the human cognitive trajectory of ``identifying ports, recognizing devices, and inferring connections''. Secondly, we customize a multi-level reward that maps reasoning and answers into graph structures and node sets, jointly optimizing logical consistency and netlist accuracy via graph similarity and set matching. Thirdly, TGRL contains a reflective learning mechanism for low-scoring samples, which corrects the reasoning trajectory through reference answers as hints, avoiding local optima caused by sparse reward signals or erroneous reasoning paths. Moreover, we construct a circuit image-netlist reasoning dataset with 3,100 samples, offering step-by-step annotations for converting circuit images to netlists. Extensive experiments demonstrate that Circuit-Think achieves SOTA netlist accuracy and significantly improves the accuracy of downstream tasks.

Chengze Jiang, Minjing Dong, Xinli Shi, Jie Gui

Vision-language pre-training models (VLPs) demonstrate strong multimodal understanding and zero-shot generalization, yet remain vulnerable to adversarial examples, raising concerns about their reliability. Recent work, Test-Time Counterattack (TTC), improves robustness by generating perturbations that maximize the embedding deviation of adversarial inputs using PGD, pushing them away from their adversarial representations. However, due to the fundamental difference in optimization objectives between adversarial attacks and counterattacks, generating counterattacks solely based on gradients with respect to the adversarial input confines the search to a narrow space. As a result, the counterattacks could overfit limited adversarial patterns and lack the diversity to fully neutralize a broad range of perturbations. In this work, we argue that enhancing the diversity and coverage of counterattacks is crucial to improving adversarial robustness in test-time defense. Accordingly, we propose Directional Orthogonal Counterattack (DOC), which augments counterattack optimization by incorporating orthogonal gradient directions and momentum-based updates. This design expands the exploration of the counterattack space and increases the diversity of perturbations, which facilitates the discovery of more generalizable counterattacks and ultimately improves the ability to neutralize adversarial perturbations. Meanwhile, we present a directional sensitivity score based on averaged cosine similarity to boost DOC by improving example discrimination and adaptively modulating the counterattack strength. Extensive experiments on 16 datasets demonstrate that DOC improves adversarial robustness under various attacks while maintaining competitive clean accuracy.

Alvi Md Ishmam, Najibul Haque Sarker, Zaber Ibn Abdul Hakim, Chris Thomas

Multimodal Large Language Models (MLLMs) have achieved remarkable performance across vision-language tasks. Recent advancements allow these models to process multiple images as inputs. However, the vulnerabilities of multi-image MLLMs remain unexplored. Existing adversarial attacks focus on single-image settings and often assume a white-box threat model which is impractical in many real-world scenarios. This paper introduces LAMP, a black-box method for learning UAPs targeting multi-image MLLMs. LAMP applies an attention-based constraint that which prevents the model from effectively aggregating information across images. LAMP also introduces a novel cross-image contagious constraint that forces perturbed tokens to influence clean tokens to spread adversarial effects without requiring all inputs to be modified. Additionally, an index-attention suppression loss creates a robust position invariant attack. Experimental results show that LAMP outperforms SOTA baselines and achieves the highest attack success rates across multiple vision-language tasks.

Nakamasa Inoue, Kanoko Goto, Masanari Oi, Martyna Gruszka, Mahiro Ukai, Takumi Hirose, Yusuke Sekikawa

Large vision-language models (LVLMs) have shown impressive performance across a broad range of multimodal tasks. However, robust image caption evaluation using LVLMs remains challenging, particularly under domain-shift scenarios. To address this issue, we introduce the Distribution-Aware Score Decoder (DISCODE), a novel finetuning-free method that generates robust evaluation scores better aligned with human judgments across diverse domains. The core idea behind DISCODE lies in its test-time adaptive evaluation approach, which introduces the Adaptive Test-Time (ATT) loss, leveraging a Gaussian prior distribution to improve robustness in evaluation score estimation. This loss is efficiently minimized at test time using an analytical solution that we derive. Furthermore, we introduce the Multi-domain Caption Evaluation (MCEval) benchmark, a new image captioning evaluation benchmark covering six distinct domains, designed to assess the robustness of evaluation metrics. In our experiments, we demonstrate that DISCODE achieves state-of-the-art performance as a reference-free evaluation metric across MCEval and four representative existing benchmarks.

Yuyang Huang, Tianzuo Luo, Hengyuan Guo, Yuren Zhang

Vision-Language Models (VLMs) have advanced multimodal understanding, yet they remain susceptible to adversarial attacks. Among various strategies, transfer-based attacks are notably effective, especially in black-box scenarios. The dominant approach within this paradigm leverages generative models to create image targets from text, consistently outperforming text-only methods. However, this approach suffers from a fundamental limitation: generative models introduce visual features irrelevant or even detrimental to textual semantics, misguiding optimization. To investigate this limitation, we conduct comprehensive analysis revealing two critical findings. First, optimal attack directions lie in synergistic spaces between image and text gradients, demonstrating that text provides complementary information. Second, widespread gradient conflicts occur when combining modalities, where image-target gradients oppose text-target directions. This conflict provides direct evidence that extraneous visual information actively harms optimization, driving it away from intended textual objectives. Based on these insights, we propose Text-Guided Gradient Refinement (TGGR), a novel framework that employs a conflict-aware projection mechanism to resolve this conflict. TGGR preserves the beneficial characteristics of image targets by decomposing the image gradient and surgically removing components that oppose the textual guidance. Extensive experiments on models such as LLaVA and GPT-4o demonstrate that TGGR substantially improves attack success rates. Specifically, on GPT-4o, TGGR yields an improvement of up to 14% over state-of-the-art methods, achieving 96% attack success rate. Our work offers a principled framework for developing more synergistic and effective adversarial strategies against VLMs.

Weiquan Huang, Aoqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Usman Naseem, Chunyu Wang, Qi Dai, Xiyang Dai, Dongdong Chen 等

CLIP is a seminal multimodal model that maps images and text into a shared representation space by contrastive learning on billions of image–caption pairs. Inspired by the rapid progress of large language models (LLMs), we investigate how the superior linguistic understanding and broad world knowledge of LLMs can further strengthen CLIP—particularly in handling long, complex captions. We introduce an efficient fine-tuning framework that embeds an LLM into a pretrained CLIP while incurring almost the same training cost as regular CLIP fine-tuning. Our method first “embedding-izes” the LLM for the CLIP setting, then couples it to the pretrained CLIP vision encoder through a lightweight adaptor trained on only a few million image–caption pairs. With this strategy we achieve large performance gains—without large-scale retraining—over state-of-the-art CLIP variants such as EVA02 and SigLIP-2. The LLM-enhanced CLIP delivers consistent improvements across a wide spectrum of downstream tasks, including linear-probe classification, zero-shot image–text retrieval with both short and long captions (in English and other languages), zero-shot/supervised image segmentation, object detection, and used as tokenizer for multimodal large-model benchmarks.

Yuhang Han, Xuyang Liu, Zihan Zhang, Pengxiang Ding, Junjie Chen, Honggang Chen, Donglin Wang, Qingsen Yan, Siteng Huang

The quadratic complexity of Multimodal Large Language Models (MLLMs) with respect to context length poses significant computational and memory challenges, hindering their real-world deployment. In the paper, we devise a ''filter-correlate-compress'' framework to accelerate the MLLM by systematically optimizing multimodal context length during prefilling. The framework first implements FiCoCo-V, a training-free method operating within the vision encoder. It employs a redundancy-based token discard mechanism that uses a novel integrated metric to accurately filter out redundant visual tokens. To mitigate information loss, the framework introduces a correlation-based information recycling mechanism that allows preserved tokens to selectively recycle information from correlated discarded tokens with a self-preserving compression, thereby preventing the dilution of their own core content. The framework's FiCoCo-L variant further leverages task-aware textual priors to perform token reduction directly within the LLM decoder. Extensive experiments demonstrate that the FiCoCo series effectively accelerates a range of MLLMs, achieves up to 14.7× FLOPs reduction with 93.6% performance retention. Our methods consistently outperform state-of-the-art training-free approaches, showcasing effectiveness and generalizability across model architectures, sizes, and tasks without requiring retraining.

Qi Guo, Xiaojun Jia, Shanmin Pang, Simeng Qin, Lin Wang, Ju Jia, Yang Liu, Qing Guo

Multimodal Large Language Models (MLLMs) are becoming integral to autonomous driving (AD) systems due to their strong vision-language reasoning capabilities. However, MLLMs are vulnerable to adversarial attacks—particularly adversarial patch attacks—which can pose serious threats in real-world scenarios. Existing patch-based attack methods are primarily designed for object detection models. Due to the more complex architectures and strong reasoning capabilities of MLLMs, these approaches perform poorly when transferred to MLLM-based systems. To address these limitations, we propose PhysPatch, a physically realizable and transferable adversarial patch framework tailored for MLLM-based AD systems. PhysPatch jointly optimizes patch location, shape, and content to enhance attack effectiveness and real-world applicability. It introduces a semantic-based mask initialization strategy for realistic placement, an SVD-based local alignment loss with patch-guided crop-resize to improve transferability, and a potential field-based mask refinement method. Extensive experiments across open-source, commercial, and reasoning-capable MLLMs demonstrate that PhysPatch significantly outperforms state-of-the-art (SOTA) methods in steering MLLM-based AD systems toward target-aligned perception and planning outputs. Moreover, PhysPatch consistently places adversarial patches in physically feasible regions of AD scenes, ensuring strong real-world applicability and deployability.

Xiang Gu, Liming Lu, Xu Zheng, Anan Du, Yongbin Zhou, Shuchao Pang

Adversarial attacks pose a significant threat to learning-based 3D point cloud models, critically undermining their reliability in security-sensitive applications. Existing defense methods often suffer from (1) high computational overhead and (2) poor generalization ability across diverse attack types. To bridge these gaps, we propose a novel yet efficient teacher-student framework, namely Multimodal Robust Prompt Distillation (MRPD) for distilling robust 3D point cloud model. It learns lightweight prompts by aligning student point cloud model's features with robust embeddings from three distinct teachers: a vision model processing depth projections, a high-performance 3D model, and a text encoder. To ensure a reliable knowledge transfer, this distillation is guided by a confidence-gated mechanism which dynamically balances the contribution of all input modalities. Notably, since the distillation is all during the training stage, there is no additional computational cost at inference. Extensive experiments demonstrate that MRPD substantially outperforms state-of-the-art defense methods against a wide range of white-box and black-box attacks, while even achieving better performance on clean data. Our work presents a new, practical paradigm for building robust 3D vision systems by efficiently harnessing multimodal knowledge.

Fuqiang Gu, Yuanke Li, Xianlei Long, Kangping Ji, Chao Chen, Qingyi Gu, Zhenliang Ni

Semantic segmentation is a fundamental task in computer vision with wide-ranging applications, including autonomous driving and robotics. While RGB-based methods have achieved strong performance with CNNs and Transformers, their effectiveness degrades under fast motion, low-light, or high dynamic range conditions due to limitations of frame cameras. Event cameras offer complementary advantages such as high temporal resolution and low latency, yet lack color and texture, making them insufficient on their own. To address this, recent research has explored multimodal fusion of RGB and event data; however, many existing approaches are computationally expensive and focus primarily on spatial fusion, neglecting the temporal dynamics inherent in event streams. In this work, we propose MambaSeg, a novel dual-branch semantic segmentation framework that employs parallel Mamba encoders to efficiently model RGB images and event streams. To reduce cross-modal ambiguity, we introduce the Dual-Dimensional Interaction Module (DDIM), comprising a Cross-Spatial Interaction Module (CSIM) and a Cross-Temporal Interaction Module (CTIM), which jointly perform fine-grained fusion along both spatial and temporal dimensions. This design improves cross-modal alignment, reduces ambiguity, and leverages the complementary properties of each modality. Extensive experiments on the DDD17 and DSEC datasets demonstrate that MambaSeg achieves state-of-the-art segmentation performance while significantly reducing computational cost, showcasing its promise for efficient, scalable, and robust multimodal perception.

Qinghao Gao, Jiahui Qu, Wenqian Dong

Multimodal remote sensing image joint classification has achieved significant progress. However, existing methods primarily focus on designing modality-specific networks, lacking adaptive generalization capabilities in diverse and dynamic modality combinations encountered in real-world scenarios. Inspired by the generalization capabilities of visual foundation model in downstream tasks, we propose a unified Text-guided Arbitrary Modalitiy Prompting (T-APT) framework, which leverages complementary fused features to drive the foundation model and employs text-guided modality-specific prior knowledge as cross-modal prompts to fine-tune a pretrained Vision Transformer (ViT) model. Specifically, a Mamba-Based Arbitrary Modal-Focused Feature Capture (MAMF-FC) module is designed to extract complementary joint features and modality-specific prior knowledge from arbitrary modalities through a shared-specific scanning encoder-decoder architecture. Subsequently, a Text-Guided Modality-Aware Prompt Tuning (TMPT) module is proposed to support the adaptation of fused features to the foundation model, enabling our arbitrary remote sensing image classification task. Extensive experiments on public datasets spanning multispectral (MS), hyperspectral (HS), light detection and ranging (LiDAR), and synthetic aperture radar (SAR) modalities demonstrate that our T-APT achieves classification performance comparable to specialized networks across arbitrary modal combinations.

Qi Gao, Heng Li, Yixin Zhou, Meixuan Zhou, Jieqiong Chen, Xinyu Chai

We present VisAssist, the first large-scale video question-answering dataset with 13,413 real-world videos captured by visually impaired users, addressing a critical gap in assistive vision research. Unlike existing benchmarks relying on third-person footage, VisAssist provides authentic first-person perspectives that uniquely capture challenges in blind photography—including unconventional framing, motion artifacts, and frequent information omission. Benchmark evaluations of SOTA multimodal models reveal systematic limitations: severe deficiencies in spatial reasoning when processing dynamic first-person viewpoints, an inability to distinguish missing information from poor capture quality leading to hazardous hallucinations, and fragile text understanding especially for non-Latin scripts under suboptimal conditions. This work establishes a vital real-world benchmark and underscores the need for specialized architectures in visual assistance systems.

Peng Gao, Yujian Lee, Xiaofeng Zhang, Zailong Chen, Hui Zhang

Large Vision-Language Models (LVLMs) have achieved impressive performance across a wide range of multimodal tasks. However, they still face critical challenges in modeling long-range dependencies under the usage of Rotary Positional Encoding (ROPE). Although it can facilitate precise modeling of token positions, it induces progressive attention decay as token distance increases, especially with progressive attention decay over distant token pairs, which severely impairs the model's ability to remember global context. To alleviate this issue, we propose inference-only Three-step Decay Resilience Strategies (T-DRS), comprising (1) Semantic-Driven DRS (SD-DRS), amplifying semantically meaningful but distant signals via content-aware residuals, (2) Distance-aware Control DRS (DC-DRS), which can purify attention by smoothly modulating weights based on positional distances, suppressing noise while preserving locality, and (3) re-Reinforce Distant DRS (reRD-DRS), consolidating the remaining informative remote dependencies to maintain global coherence. Together, the T-DRS recover suppressed long-range token pairs without harming local inductive biases. Extensive experiments on Vision Question Answering (VQA) benchmarks demonstrate that T-DRS can consistently improve performance in an inference-only manner.

Yu Fei, Quan Deng, Shengeng Tang, Li Yuehua, Lechao Cheng

Open-world 3D scene understanding is fundamentally challenging for vision and robotics, due to the constraints of closed-vocabulary supervision and static annotations. To address this, we propose a unified framework for Open-World 3D Scene Graph Generation with Retrieval-Augmented Reasoning, which enables generalizable and interactive 3D scene understanding. Our method integrates vision-language models with retrieval-based reasoning to support multimodal exploration and language-guided interaction. The framework comprises two key components: (1) a dynamic scene graph generation module that detects objects and infers semantic relationships without fixed label sets, and (2) a retrieval-augmented reasoning pipeline that encodes scene graphs into a vector database to support text/image-conditioned queries. We evaluate our method on 3DSSG and Replica benchmarks across four tasks—scene question answering, visual grounding, instance retrieval, and task planning—demonstrating robust generalization and superior performance in diverse environments. Our results highlight the effectiveness of combining open-vocabulary perception with retrieval-based reasoning for scalable 3D scene understanding.

Nastaran Darabi, Devashri Naik, Sina Tayebati, Dinithi Jayasuriya, Ranganath Krishnan, Amit Ranjan Trivedi

Large Language Models (LLMs) and Vision-Language Models (VLMs) remain highly vulnerable to adversarial attacks despite widespread adoption. Existing defenses typically require retraining, rely on heuristics, or fail under adaptive and out-of-distribution (OOD) conditions. We introduce EigenShield, a principled, inference-time, architecture-agnostic defense that leverages Random Matrix Theory (RMT) to suppress adversarial noise in high-dimensional embeddings. EigenShield uses spiked covariance modeling and a Robustness-based Nonconformity Score (RbNS) with quantile thresholding to isolate and preserve causal eigenvectors, filtering out adversarial components without model access or adversarial training. We develop a theoretical framework establishing conditions for asymptotic noise suppression and demonstrate effectiveness in both unimodal and multimodal settings. Empirically, EigenShield consistently improves robustness across threat models, reducing attack success rates (ASR) by up to 48% over state-of-the-art defenses, including adversarial training, UNIGUARD, CIDER, and input transformations. On jailbreak attacks, EigenShield lowers LLM ASR by up to 92.9% relative to undefended models. Under multimodal adversarial attacks, it reduces VLM ASR by up to 76.5%. Against adaptive attacks on LLMs, it achieves ASR reductions of up to 77.7%. In OOD settings, EigenShield maintains strong performance, reducing ASR by up to 88.4% for LLMs and 80.4% for VLMs.

Dahyun Chung, Donghyun Shin, Yujin Sung, Seunggi Moon, Jinwoo Jeon, Byung-Jun Lee

Contrastive Language–Image Pre-training (CLIP) has demonstrated strong generalization across a wide range of visual tasks by leveraging large-scale English–image pairs. However, its extension to low-resource languages remains limited due to the scarcity of high-quality multilingual image–text data. Existing multilingual vision–language models exhibit consistently low retrieval performance in underrepresented languages—including Czech, Finnish, Croatian, Hungarian, Romanian—on the Crossmodal-3600 (XM3600) benchmark. To address this, we propose a lightweight and data-efficient framework for multilingual vision–language alignment. Our approach requires no image–text pairs or text-text pairs and freezes both the pretrained image encoder and multilingual text encoder during training. Only a compact 1.7M-parameter projection module is trained, using a contrastive loss over English representations as semantic anchors. This minimal training setup enables robust multilingual alignment even for languages with limited supervision. Extensive evaluation across multiple multilingual retrieval benchmarks confirms the effectiveness of our method, showing significant gains in five underrepresented languages where existing models typically underperform. These findings highlight the effectiveness of our pivot-based, parameter-efficient alignment strategy for inclusive multimodal learning.

Meng Chu, Yukang Chen, Haokun Gui, Shaozuo Yu, Yi Wang, Jiaya Jia

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 comprehensive travel assistance. Our work addresses the fundamental challenge of developing practical AI travel assistants through three key contributions: (1) TravelQA, a novel dataset of 265k question-answer pairs combining 160k text QA from authentic travel sources, 100k vision-language QA featuring maps and location imagery, and 5k expert-annotated Chain-of-Thought reasoning examples; (2) Travel-CoT, a structured reasoning framework that decomposes travel queries into spatial, temporal, and practical dimensions, improving answer accuracy by 10.8% while providing interpretable decision paths; and (3) an interactive agent system validated through extensive user studies. Through fine-tuning experiments on state-of-the-art vision-language models (LLaVA, Qwen-VL, Shikra), we achieve 6.2-9.4% base improvements, further enhanced by Travel-CoT reasoning. Our model demonstrates superior capabilities in contextual travel recommendations, map interpretation, and scene understanding while providing practical information such as operating hours and cultural insights. User studies with 500 participants show TraveLLaMA achieves a System Usability Scale score of 82.5, significantly outperforming general-purpose models and establishing new standards for multimodal travel assistance systems.

Junyu Cheng, Zhibiao Liang, Yidong Chen, Shuangyin Li

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in multimodal understanding and generation by integrating visual and textual data. However, these models frequently exhibit object hallucination problems: generating outputs that are inconsistent with the input image. Existing improved methods for mitigating hallucinations still suffer from two key limitations: dynamic approaches based on logits or attention mechanisms risk suppressing valuable linguistic priors, whereas static methods that employ fixed intervention vectors lack the flexibility to adapt to diverse images and questions. To address these issues, we propose RFI (Rectified Flow Intervention), a novel approach that harnesses the linear trajectory design of rectified flow for input-specific adaptation and employs gradient correction to ensure coherent generation, effectively combining the adaptability of dynamic methods with the stability of static ones. RFI dynamically predicts latent-space intervention vectors while requiring only a single forward pass in LVLMs per question, achieving computational efficiency (1.09x latency overhead for 100 new tokens). Extensive experiments show RFI significantly reduces hallucinations, achieving superior performance compared to existing advanced methods, highlighting its effectiveness as a lightweight plug-and-play method for reducing LVLM's hallucination in practical applications.

Jingchao Cao, Guo An, Feng Gao, Ke Gu, Yutao Liu

Despite the rapid progress of multimodal large language models (MLLMs), their capacity for low-level visual perception in underwater environments remains underexplored. To address this gap, we present UQ-Bench, the first systematically designed benchmark for evaluating the ability of MLLMs to perceive and assess underwater image quality at the low-level visual attribute level. UQ-Bench comprises three components: (1) UW-Perception, a dataset of 3,000 underwater images paired with targeted questions on key degradations such as color cast, blur, contrast, and exposure, covering both global and local perceptual dimensions; (2) UW-Describe, a dataset of 500 images with expert-annotated gold-standard descriptions for assessing the accuracy of model-generated text; and (3) UW-Eval, an evaluation protocol employing human mean opinion scores (MOS) for quantitative quality assessment. To ensure rigorous and reproducible benchmarking, we propose a GPT-assisted evaluation framework that aligns model outputs with expert references and enables fine-grained analysis of distortion perception. Experimental results demonstrate that while MLLMs exhibit preliminary competence in underwater low-level visual tasks, they still fall short in capturing subtle degradations and achieving human-level consistency, highlighting the need for further advances in foundation models for marine vision.