论文检索

Yulong Li 0002, Yuxuan Zhang, Rui Chen, Feilong Tang, Zhixiang Lu, Ming Hu, Jianghao Wu 0001, Haochen Xue, Mian Zhou, Chong Li 等

Artificial intelligence will not achieve genuine empathy until models can reason about the causes of human emotions rather than only label them. Current datasets fail to support this objective, as existing emotional causality datasets primarily focus on textual modalities, lack non-verbal information such as speech and facial expressions, feature relatively short dialogue lengths, and limit research on long-term emotional evolution. Existing annotations concentrate on stimulus-response patterns and lack cross-temporal emotional causal chain annotations, failing to reveal how early events accumulate and ultimately trigger emotional changes. In this work, we introduce Genesis, the first multimodal dialogue dataset supporting long-term emotional causality analysis, which Genesis contains 1,000 dialogues averaging 208 turns each, spanning debate, family, educational, and social scenarios. Through two-layer annotation system: proximal cause identification and long-term causal chain tracking, Genesis labels complex emotional phenomena including cross-modal inconsistencies and long-distance causal dependencies. Our evaluation of 20 mainstream multimodal models reveals limitations in current approaches for long-term emotional causality. We propose Empathica as an evaluation baseline, employing a Recognition-Memory-Attribution architecture that integrates dynamic sliding windows and event aggregation mechanisms to address multimodal emotional causality modeling challenges. Empathica outperforms text-based models GPT-o1, and multimodal model Gemini 1.5 Pro and GPT-4o across all evaluation metrics.

Cheng Zhang, Hongxia Xie, Bin Wen 0001, Songhan Zuo, Ruoxuan Zhang, Wen-Huang Cheng

With the rapid advancement of diffusion models, text-to-image generation has achieved significant progress in image resolution, detail fidelity, and semantic alignment, particularly with models like Stable Diffusion 3.5, Stable Diffusion XL, and FLUX.1. However, generating emotionally expressive and abstract artistic images remains a major challenge, largely due to the lack of large-scale, fine-grained emotional datasets. To address this gap, we present the EmoArt Dataset-one of the most comprehensive emotion-annotated art datasets to date. It contains 132,664 artworks across 56 painting styles (e.g., Impressionism, Expressionism, Abstract Art), offering rich stylistic and cultural diversity. Each image includes structured annotations: objective scene descriptions, five key visual attributes (brushwork, composition, color, line, light), binary arousal-valence labels, twelve emotion categories, and potential art therapy effects. Using EmoArt, we systematically evaluate popular text-to-image diffusion models for their ability to generate emotionally aligned images from text. Our work provides essential data and benchmarks for emotion-driven image synthesis and aims to advance fields such as affective computing, multimodal learning, and computational art, enabling applications in art therapy and creative design. The dataset and more details can be accessed via the following link: https://zhiliangzhang.github.io/EmoArt-130k/

Quang-Trung Truong, Yuk-Kwan Wong, Vo Hoang Kim Tuyen Dang, Rinaldi Gotama, Duc Thanh Nguyen, Sai-Kit Yeung

Marine videos present significant challenges for video understanding due to the dynamics of marine objects and the surrounding environment, camera motion, and the complexity of underwater scenes. Existing video captioning datasets, typically focused on generic or human-centric domains, often fail to generalize to the complexities of the marine environment and gain insights about marine life. To address these limitations, we propose a two-stage marine object-oriented video captioning pipeline. We introduce a comprehensive video understanding benchmark that leverages the triplets of video, text, and segmentation masks to facilitate visual grounding and captioning, leading to improved marine video understanding and analysis, and marine video generation. Additionally, we highlight the effectiveness of video splitting in order to detect salient object transitions in scene changes, which significantly enrich the semantics of captioning content. Our dataset and code have been released at https://msc.hkustvgd.com.

Tao Feng, Tingfa Xu, Haolin Qin, Tianhao Li, Shuaihao Han, Xuyang Zou, Zhan Lv, Jianan Li 0001

Visual object tracking in real-world scenarios presents numerous challenges including occlusion, interference from similar objects and complex backgrounds - all of which limit the effectiveness of RGB-based trackers. Multispectral imagery, which captures pixel-level spectral reflectance, enhances target discriminability. However, the availability of multispectral tracking datasets remains limited. To bridge this gap, we introduce MSITrack, the largest and most diverse multispectral single object tracking dataset to date. MSITrack offers the following key features: (i) More Challenging Attributes - including interference from similar objects and similarity in color and texture between targets and backgrounds in natural scenarios, along with a wide range of real-world tracking challenges; (ii) Richer and More Natural Scenes - spanning 55 object categories and 300 distinct natural scenes, MSITrack far exceeds the scope of existing benchmarks. Many of these scenes and categories are introduced to the multispectral tracking domain for the first time; (iii) Larger Scale - 300 videos comprising over 129k frames of multispectral imagery. To ensure annotation precision, each frame has undergone meticulous processing, manual labeling and multi-stage verification. Extensive evaluations using representative trackers demonstrate that the multispectral data in MSITrack significantly improves performance over RGB-only baselines, highlighting its potential to drive future advancements in the field. The MSITrack dataset is publicly available at: https://github.com/Fengtao191/MSITrack.

Qiyuan Guan, Qianfeng Yang, Xiang Chen 0015, Tianyu Song 0003, Guiyue Jin, Jiyu Jin

Existing all-in-one image restoration approaches, which aim to handle multiple weather degradations within a single framework, are predominantly trained and evaluated using mixed single-weather synthetic datasets. However, these datasets often differ significantly in resolution, style, and domain characteristics, leading to substantial domain gaps that hinder the development and fair evaluation of unified models. Furthermore, the lack of a large-scale, real-world all-in-one weather restoration dataset remains a critical bottleneck in advancing this field. To address these limitations, we present a real-world all-in-one adverse weather image restoration benchmark dataset, which contains image pairs captured under various weather conditions, including rain, snow, and haze, as well as diverse outdoor scenes and illumination settings. The resulting dataset provides precisely aligned degraded and clean images, enabling supervised learning and rigorous evaluation. We conduct comprehensive experiments by benchmarking a variety of task-specific, task-general, and all-in-one restoration methods on our dataset. Our dataset offers a valuable foundation for advancing robust and practical all-in-one image restoration in real-world scenarios. The dataset has been publicly released and is available at https://github.com/guanqiyuan/WeatherBench.

Hengrui Lou, Zunlei Feng, Jinsong Geng, Erteng Liu, Lechao Cheng, Jie Lei 0002, Jie Song 0011, Mingli Song, Yijun Bei

With the rapid advancement of AIGC technology, realistic fake facial images and videos that deceive human perception are now possible. Consequently, numerous face forgery detection techniques have been proposed. However, evaluating their effectiveness and generalizability remains a significant challenge. To address this, we introduce DeepFaceGen, a large-scale benchmark for quantitatively assessing face forgery detection performance and supporting iterative advancements in the field. DeepFaceGen comprises 776,990 real face images/videos and 773,812 forged face samples generated using 35 mainstream face generation techniques. During its construction, we prioritized content diversity, ethnic fairness, and comprehensive labeling to ensure its versatility and usability. DeepFaceGen is then used to evaluate 20 leading face forgery detection methods from multiple perspectives. Through extensive analysis, we present key insights and propose directions for future research. The code and dataset for DeepFaceGen are available at https://github.com/HengruiLou/DeepFaceGen

Wentao Liu, Qianjun Pan, Yi Zhang, Zhuo Liu, Ji Wu, Jie Zhou 0015, Aimin Zhou, Qin Chen 0001, Bo Jiang 0016, Liang He 0001

Large language models (LLMs) have obtained promising results in mathematical reasoning, a foundational human intelligence skill. Most previous studies focus on improving or measuring the performance of LLMs via textual math datasets (e.g., MATH, GSM8K). In this paper, we release a Chinese multimodal math (CMM-Math) dataset, including benchmark and training parts, to evaluate and enhance the mathematical reasoning of LMMs. CMM-Math contains over 28,000 high-quality samples, featuring a variety of problem types (e.g., choice, fill-in-the-blank, analysis) with detailed solutions across 12 grade levels from elementary to high school in China. The problem may contain multiple images, and the visual context may be present in the questions or opinions, which makes this dataset more challenging. Our comprehensive analysis reveals that state-of-the-art LMMs on the CMM-Math face challenges, emphasizing the necessity for further improvements in LMM development. We also propose a Multimodal Mathematical LMM (Math-LMM) to handle the problems with mixed input of multiple images and text segments. The Math-LMM is trained using three stages: foundational pre-training, foundational fine-tuning, and mathematical fine-tuning. The extensive experiments indicate that our model effectively improves math reasoning performance by comparing it with the SOTA LMMs over three multimodal mathematical datasets. We release the datasets on GitHub (https://github.com/ECNU-ICALK/EduChat-Math) and Huggingface (https://huggingface.co/datasets/ecnu-icalk/cmm-math).

Abel Yu Hao Chai, Kelly Li Zhen Jee, Sue Han Lee, Fei Siang Tay, Jules Vandeputte, Hervé Goëau, Pierre Bonnet, Alexis Joly

Deep learning models have emerged as a promising alternative to conventional approaches for plant disease identification, a critical challenge in agricultural production. However, the existing plant disease datasets are insufficient to address the complexities of real-world agricultural scenarios, such as multi crop disease, unseen, few-shot, and domain shift adaptation. Additionally, the lack of standardized evaluation protocols and benchmark datasets hinders the fair evaluation of models against these challenges. To bridge this gap, we introduce Deep-Plant-Disease, the largest and most diverse dataset with novel text data designed to enhance model generalization in multi crop disease identification. We revisit and reformulate the task by establishing a standardized evaluation framework that supports consistent benchmarking and guides future research. Through experiments, we further validate the robustness and adaptability of models trained on our dataset, highlighting their effective transferability to real-world agricultural challenges.

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.

Feng-Kai Huang, Hong-Wei Xu, Chu-Chuan Lee, Hong-Yi Tu, Hong-Han Shuai, Wen-Huang Cheng

Long-term multi-animal tracking in densely group-housed agricultural settings is critical for automated behavior monitoring and early anomaly detection in precision livestock farming. However, it poses significant challenges due to persistent occlusions from feeders and water dispensers, high inter-individual appearance similarity, and drastic visual changes across day and night cycles. Existing multi-object tracking datasets rarely capture the combined difficulty of these real-world conditions. To address this, we introduce OinkTrack, a large-scale benchmark for continuous multi-pig tracking in commercial farm environments. The dataset comprises over five hours of annotated video across sixteen sequences, covering day, night, night-to-day, and day-to-night transitions. Each sequence ranges from one minute to one hour, featuring an average of thirty-six pigs per frame. In total, OinkTrack provides 573,700 bounding boxes linked to 574 consistent pig identities. It enables detailed behavior analysis under varying lighting and crowding conditions. We describe the data collection and annotation process, present statistical insights into tracking difficulty, and benchmark 11 state-of-the-art tracking methods. OinkTrack provides a robust foundation for developing long-term tracking models and supports downstream applications such as individual activity profiling and early detection of abnormal behavior in real-world, high-density animal populations. The complete dataset and supplementary materials are publicly accessible at https://leohuang0511.github.io/oinktrack-page.

Hongru Xiao, Xiang Li 0064, Duyi Pan, Longfei Zhang, ZhixueSong ZhixueSong, Jiale Han 0001, Songning Lai, Wenshuo Chen, Jing Tang, Benyou Wang

Recent advancements in Audio Language Models (ALMs) have led to significant improvements in speech-related tasks. However, their capacity for profound metaphorical reasoning, especially when derived from audio-specific cues, has yet to be thoroughly investigated. To address this gap, we introduce Unspoken, a bilingual (Chinese-English) question answering benchmark designed to assess ALMs' comprehension of non-literal, metaphor-rich audio. Unlike prior text-centric evaluations, Unspoken emphasizes prosody, phonetic ambiguity, emotional inflection, and other nuanced acoustic features critical to metaphor understanding but often lost in transcription. We construct a high-quality dataset of 2,764 manually curated and validated QA pairs, spanning three reasoning dimensions: semantic, acoustic, and contextual, and covering six common types of metaphors. Evaluation across 23 mainstream ALMs reveals a substantial performance gap: the best model achieves only 69.5% accuracy, significantly below the human average of 81.1%. By analyzing the error patterns, we identify five key failure modes that reveal fundamental limitations in current models' reasoning capabilities. Unspoken not only sets a new standard for evaluating metaphorical reasoning in audio but also pioneers a novel research direction that moves beyond transcription-based assessments. Grounding metaphor understanding in authentic human communication scenarios offers deep insight for developing more cognitively capable ALMs. The data and codes are available at https://github.com/Hongru0306/UNSPOKEN.

Nidham Tekaya, Manuela Waldner, Matthias Zeppelzauer

Large-scale vision-language models (VLMs) such as CLIP have gained popularity for their generalizable and expressive multimodal representations. By leveraging large-scale training data with diverse textual metadata, VLMs acquire open-vocabulary capabilities, solving tasks beyond their training scope. This paper investigates the temporal awareness of VLMs, assessing their ability to position visual content in time. We introduce TIME10k, a benchmark dataset of over 10,000 images with temporal ground truth, and evaluate the time-awareness of 37 VLMs by a novel methodology. Our investigation reveals that temporal information is structured along a low-dimensional, non-linear manifold in the VLM embedding space. Based on this insight, we propose methods to derive an explicit ''timeline'' representation from the embedding space. These representations model time and its chronological progression and thereby facilitate temporal reasoning tasks. Our timeline approaches achieve competitive to superior accuracy compared to a prompt-based baseline while being computationally efficient. All code and data are available at https://tekayanidham.github.io/timeline-page/.

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.

Zihao Wang, Ruibin Yuan, Ziqi Geng, Hengjia Li, Xingwei Qu, Xinyi Li, Songye Chen, Haoying Fu, Roger B. Dannenberg, Kejun Zhang

Automated singing assessment is crucial for education, entertainment, and talent discovery. However, existing systems are hindered by two fundamental limitations: first, their reliance on reference tracks (e.g., the original song), which stifles creative expression, and second, their simplification of complex vocal performances into a single, often non-diagnostic score based on pitch and rhythm. This paradigm fails to capture the nuanced, multifaceted attributes that define expert-level singing. Echoing the recent shift in other AI domains from discriminative to descriptive evaluation, we advocate for a new paradigm in singing assessment. This paper aims to build a complete ecosystem for reference-free, multi-dimensional, and descriptive singing assessment. First, we construct Sing-MD, a large-scale, multi-dimensional singing dataset annotated by experts across four core dimensions: breath control, timbre quality, emotional expression, and vocal technique. Analysis of this dataset reveals a key finding: significant annotation inconsistencies among experts, which challenges the validity of traditional accuracy-based evaluation metrics. Second, standard Multimodal Large Language Models (MLLMs) are unable to analyze full-length songs on resource-constrained, consumer-grade hardware due to memory limitations. This challenge leads to a ''human label-audio input mismatch'' problem and results in poor performance. To address this issue, we designed VocalVerse, an efficient hybrid architecture. It leverages a lightweight acoustic encoder and specialized modules to process the entire song, thereby learning global performance features, modeling long-term dependencies, and ultimately overcoming this limitation. Third, to address the shortcomings of automated metrics, we establish a new evaluation benchmark-H-TPR (Human-in-the-loop Tiered Perceptual Ranking)-which evaluates a model's ability to generate perceptually valid performance rankings, rather than predicting a noisy ''ground-truth'' score. Our comprehensive experiments show that on the H-TPR benchmark, our VocalVerse framework can effectively learn and distinguish singing quality across different dimensions, thereby creating perceptually valid quality rankings and significantly outperforming existing baselines. Furthermore, our framework for multi-dimensional scoring and descriptive feedback generation has been successfully commercialized and deployed at scale, demonstrating its significant real-world impact and practical value.

Weiwu Pang, Rajrup Ghosh, Jiawei Yang 0006, Ziyu Wei, Branden Leong, Yue Wang, Ramesh Govindan

Outdoor AR applications on mobile devices need accurate estimates for the pose of the device. In this paper, we develop SplatPose, a novel pose estimation technique that uses a data-driven 3D modeling technique called Gaussian Splatting. SplatPose uses a trained Gaussian Splatting model to render an image at an estimated device location, then matches features with the camera image to estimate pose. % Because this matching can be fast, SplatPose can, in theory, estimate pose entirely on a mobile device, while existing approaches cannot. To this end, SplatPose trains Gaussian Splatting models to be robust to appearance changes, thereby improving accuracy. It also incorporates a novel fast renderer to improve rendering speed. Using an AR pose estimation benchmark dataset, we show that SplatPose outperforms the state-of-the-art in terms of accuracy, and is up to an order of magnitude faster on a mobile device.

Dong Li, Yichen Niu, Ying Ai, Xiang Zou, Biqing Qi, Jianxing Liu

Large language models (LLMs) have demonstrated strong performance in natural language generation but remain limited in knowle- dge-intensive tasks due to outdated or incomplete internal knowledge. Retrieval-Augmented Generation (RAG) addresses this by incorporating external retrieval, with GraphRAG further enhancing performance through structured knowledge graphs and multi-hop reasoning. However, existing GraphRAG methods largely ignore the temporal dynamics of knowledge, leading to issues such as temporal ambiguity, time-insensitive retrieval, and semantic redundancy. To overcome these limitations, we propose Temporal GraphRAG (T-GRAG), a dynamic, temporally-aware RAG framework that models the evolution of knowledge over time. T-GRAG consists of five key components: (1) a Temporal Knowledge Graph Generator that creates time-stamped, evolving graph structures; (2) a Temporal Query Decomposition mechanism that breaks complex temporal queries into manageable sub-queries; (3) a Three-layer Interactive Retriever that progressively filters and refines retrieval across temporal subgraphs; (4) a Source Text Extractor to mitigate noise; and (5) a LLM-based Generator that synthesizes contextually and temporally accurate responses. We also introduce Time-LongQA, a novel benchmark dataset based on real-world corporate annual reports, designed to test temporal reasoning across evolving knowledge. Extensive experiments show that T-GRAG significantly outperforms prior RAG and GraphRAG baselines in both retrieval accuracy and response relevance under temporal constraints, highlighting the necessity of modeling knowledge evolution for robust long-text question answering. Our code is publicly available on the T-GRAG https://github.com/Arvin0313/T-GRAG.git

Hoan My Tran, Damien Lolive, Aghilas Sini, Arnaud Delhay, Pierre-François Marteau, David Guennec

Recent advancements in generative AI, particularly in speech synthesis, have enabled the generation of highly natural-sounding synthetic speech that closely mimics human voices. While these innovations hold promise for applications like assistive technologies, they also pose significant risks, including misuse for fraudulent activities, identity theft, and security threats. Current research on spoofing detection countermeasures remains limited by generalization to unseen deepfake attacks and languages. To address this, we propose a gating mechanism extracting relevant feature from the speech foundation XLS-R model as a front-end feature extractor. For downstream back-end classifier, we employ Multi-kernel gated Convolution (MultiConv) to capture both local and global speech artifacts. Additionally, we introduce Centered Kernel Alignment (CKA) as a similarity metric to enforce diversity in learned features across different MultiConv layers. By integrating CKA with our gating mechanism, we hypothesize that each component helps improving the learning of distinct synthetic speech patterns. Experimental results demonstrate that our approach achieves state-of-the-art performance on in-domain benchmarks while generalizing robustly to out-of-domain datasets, including multilingual speech samples. This underscores its potential as a versatile solution for detecting evolving speech deepfake threats.

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.

Inzamamul Alam, Md Tanvir Islam, Simon S. Woo

The increasing realism of content generated by GANs and diffusion models has made deepfake detection significantly more challenging. Existing approaches often focus solely on spatial or frequency-domain features, limiting their generalization to unseen manipulations. We propose the Spectral Cross-Attentional Network (SpecXNet), a dual-domain architecture for robust deepfake detection. The core Dual-Domain Feature Coupler (DDFC) decomposes features into a local spatial branch for capturing texture-level anomalies and a global spectral branch that employs Fast Fourier Transform to model periodic inconsistencies. This dual-domain formulation allows SpecXNet to jointly exploit localized detail and global structural coherence, which are critical for distinguishing authentic from manipulated images. We also introduce the Dual Fourier Attention (DFA) module, which dynamically fuses spatial and spectral features in a content-aware manner. Built atop a modified XceptionNet backbone, we embed the DDFC and DFA modules within a separable convolution block. Extensive experiments on multiple deepfake benchmarks show that SpecXNet achieves state-of-the-art accuracy, particularly under cross-dataset and unseen manipulation scenarios, while maintaining real-time feasibility. Our results highlight the effectiveness of unified spatial-spectral learning for robust and generalizable deepfake detection. To ensure reproducibility, we release the full code on https://github.com/inzamamulDU/SpecXNet

Man Xiao, Jianbin Ye, Bo Liu 0014, Zijian Gao, Kele Xu, Xiaodong Wang 0002

Multimodal deepfakes pose growing security threats across diverse domains, driven by rapid advancements in generative models. This demands effective Multimodal Deepfake Continual Detection (MDCD) methods capable of adapting to evolving and heterogeneous deepfake techniques. However, MDCD remains underexplored, facing two major challenges: (1) modality-specific feature disparities limit the effectiveness of simple feature fusion, exacerbating the forgetting of previous forgery-relevant knowledge; and (2) newly introduced deepfake videos initially exhibit limited scale that gradually expand, causing class imbalance dominated by forged samples, undermines authentic content understanding in comming tasks. To address these issues, we propose the Analytic Synaptic Dynamic Scaling Balancer (ADanser) that adapts to modality-specific biases and class imbalance while employing a closed-form update to preserve prior multimodal deepfake knowledge in an evolving data stream. Inspired by synaptic scaling in neuroscience, ADanser introduces a modality synaptic scaling mechanism that applies modality-aware attention to extract discriminative and complementary forgery patterns, improving cross-modal knowledge retention. Additionally, a class-wise contribution balancer dynamically reweights learning signals to reduce class bias and enhance authentic video representation. Extensive experiments on benchmark multimodal deepfake datasets demonstrate that ADanser significantly outperforms state-of-the-art continual learning methods, effectively coordinating adaptation and retention in imbalanced, cross-modal scenarios.