Yanan Wang

Yanan Wang, Xuezhi Liang, Shengcai Liao

Recently, large-scale synthetic datasets are shown to be very useful for generalizable person re-identification. However, synthesized persons in existing datasets are mostly cartoon-like and in random dress collocation, which limits their performance. To address this, in this work, an automatic approach is proposed to directly clone the whole outfits from real-world person images to virtual 3D characters, such that any virtual person thus created will appear very similar to its real-world counterpart. Specifically, based on UV texture mapping, two cloning methods are designed, namely registered clothes mapping and homogeneous cloth expansion. Given clothes keypoints detected on person images and labeled on regular UV maps with clear clothes structures, registered mapping applies perspective homography to warp real-world clothes to the counterparts on the UV map. As for invisible clothes parts and irregular UV maps, homogeneous expansion segments a homogeneous area on clothes as a realistic cloth pattern or cell, and expand the cell to fill the UV map. Furthermore, a similarity-diversity expansion strategy is proposed, by clustering person images, sampling images per cluster, and cloning outfits for 3D character generation. This way, virtual persons can be scaled up densely in visual similarity to challenge model learning, and diversely in population to enrich sample distribution. Finally, by rendering the cloned characters in Unity3D scenes, a more realistic virtual dataset called ClonedPerson is created, with 5,621 identities and 887,766 images. Experimental results show that the model trained on ClonedPerson has a better generalization performance, superior to that trained on other popular real-world and synthetic person re-identification datasets. The ClonedPerson project is available at https://github.com/Yanan-Wang-cs/ClonedPerson.

Monisha Singh, Ximi Hoque, Donghuo Zeng et al.

The degree of concentration, enthusiasm, optimism, and passion displayed by individual(s) while interacting with a machine is referred to as `user engagement'. Engagement comprises of behavioral, cognitive, and affect related cues. To create engagement prediction systems that can work in real-world conditions, it is quintessential to learn from rich, diverse datasets. To this end, a large scale multi-faceted engagement in the wild dataset EngageNet is proposed. 31 hours duration data of 127 participants representing different illumination conditions are recorded. Thorough experiments are performed exploring the applicability of different features, action units, eye gaze, head pose, and MARLIN. Data from user interactions (question-answer) are analyzed to understand the relationship between effective learning and user engagement. To further validate the rich nature of the dataset, evaluation is also performed on the EngageWild dataset. The experiments show the usefulness of the proposed dataset. The code, models, and dataset link are publicly available at https://github.com/engagenet/engagenet_baselines.

Yanan Wang, Michihiro Yasunaga, Hongyu Ren et al.

Visual question answering (VQA) requires systems to perform concept-level reasoning by unifying unstructured (e.g., the context in question and answer; "QA context") and structured (e.g., knowledge graph for the QA context and scene; "concept graph") multimodal knowledge. Existing works typically combine a scene graph and a concept graph of the scene by connecting corresponding visual nodes and concept nodes, then incorporate the QA context representation to perform question answering. However, these methods only perform a unidirectional fusion from unstructured knowledge to structured knowledge, limiting their potential to capture joint reasoning over the heterogeneous modalities of knowledge. To perform more expressive reasoning, we propose VQA-GNN, a new VQA model that performs bidirectional fusion between unstructured and structured multimodal knowledge to obtain unified knowledge representations. Specifically, we inter-connect the scene graph and the concept graph through a super node that represents the QA context, and introduce a new multimodal GNN technique to perform inter-modal message passing for reasoning that mitigates representational gaps between modalities. On two challenging VQA tasks (VCR and GQA), our method outperforms strong baseline VQA methods by 3.2% on VCR (Q-AR) and 4.6% on GQA, suggesting its strength in performing concept-level reasoning. Ablation studies further demonstrate the efficacy of the bidirectional fusion and multimodal GNN method in unifying unstructured and structured multimodal knowledge.

Yanan Wang, Shuaicong Hu, Jian Liu et al.

The impressive performance of generalist large language models (LLMs) such as GPT and Claude in healthcare raises a critical question: will domain-specific medical specialist models become obsolete? We argue that the future of medical artificial intelligence (AI) lies not in building monolithic medical foundation models, nor in replacing human expertise, but in orchestrating collaboration among generalist LLMs, domain-specific specialist models, and clinicians. We propose HetMedAgent, a heterogeneous medical multi-agent framework that enables conflict-aware evidence fusion, uncertainty-based clinician intervention triggering, and adaptive threshold calibration. Experiments on three real-world clinical decision-making tasks demonstrate that the synergy between generalist LLMs and domain-specific specialist models significantly outperforms using either type of model alone, validating the irreplaceable value of specialist models in modality-specific analysis. HetMedAgent represents a shift from building medical LLMs or foundation models to multi-agent collaboration, achieving a balance between general reasoning capabilities and domain-specific precision.

Donghuo Zeng, Yanan Wang, Jianming Wu et al.

The heterogeneity gap problem is the main challenge in cross-modal retrieval. Because cross-modal data (e.g. audiovisual) have different distributions and representations that cannot be directly compared. To bridge the gap between audiovisual modalities, we learn a common subspace for them by utilizing the intrinsic correlation in the natural synchronization of audio-visual data with the aid of annotated labels. TNN-CCCA is the best audio-visual cross-modal retrieval (AV-CMR) model so far, but the model training is sensitive to hard negative samples when learning common subspace by applying triplet loss to predict the relative distance between inputs. In this paper, to reduce the interference of hard negative samples in representation learning, we propose a new AV-CMR model to optimize semantic features by directly predicting labels and then measuring the intrinsic correlation between audio-visual data using complete cross-triple loss. In particular, our model projects audio-visual features into label space by minimizing the distance between predicted label features after feature projection and ground label representations. Moreover, we adopt complete cross-triplet loss to optimize the predicted label features by leveraging the relationship between all possible similarity and dissimilarity semantic information across modalities. The extensive experimental results on two audio-visual double-checked datasets have shown an improvement of approximately 2.1% in terms of average MAP over the current state-of-the-art method TNN-CCCA for the AV-CMR task, which indicates the effectiveness of our proposed model.

Xin An, Jingyi Cai, Xiangyang Chen et al.

Addressing the challenges of fragmented task definitions and the heterogeneity of unstructured data in multimodal parsing, this paper proposes the Omni Parsing framework. This framework establishes a Unified Taxonomy covering documents, images, and audio-visual streams, introducing a progressive parsing paradigm that bridges perception and cognition. Specifically, the framework integrates three hierarchical levels: 1) Holistic Detection, which achieves precise spatial-temporal grounding of objects or events to establish a geometric baseline for perception; 2) Fine-grained Recognition, which performs symbolization (e.g., OCR/ASR) and attribute extraction on localized objects to complete structured entity parsing; and 3) Multi-level Interpreting, which constructs a reasoning chain from local semantics to global logic. A pivotal advantage of this framework is its evidence anchoring mechanism, which enforces a strict alignment between high-level semantic descriptions and low-level facts. This enables ``evidence-based'' logical induction, transforming unstructured signals into standardized knowledge that is locatable, enumerable, and traceable. Building on this foundation, we constructed a standardized dataset and released the Logics-Parsing-Omni model, which successfully converts complex audio-visual signals into machine-readable structured knowledge. Experiments demonstrate that fine-grained perception and high-level cognition are synergistic, effectively enhancing model reliability. Furthermore, to quantitatively evaluate these capabilities, we introduce OmniParsingBench. Code, models and the benchmark are released at https://github.com/alibaba/Logics-Parsing/tree/master/Logics-Parsing-Omni.

Zhaoqi Yang, Yanan Wang, Yong Ge

Sequential recommendation tasks, which aim to predict the next item a user will interact with, typically rely on models trained solely on historical data. However, in real-world scenarios, user behavior can fluctuate in the long interaction sequences, and training data may be limited to model this dynamics. To address this, Test-Time Training (TTT) offers a novel approach by using self-supervised learning during inference to dynamically update model parameters. This allows the model to adapt to new user interactions in real-time, leading to more accurate recommendations. In this paper, we propose TTT4Rec, a sequential recommendation framework that integrates TTT to better capture dynamic user behavior. By continuously updating model parameters during inference, TTT4Rec is particularly effective in scenarios where user interaction sequences are long, training data is limited, or user behavior is highly variable. We evaluate TTT4Rec on three widely-used recommendation datasets, demonstrating that it achieves performance on par with or exceeding state-of-the-art models. The codes are available at https://github.com/ZhaoqiZachYang/TTT4Rec.

Donghuo Zeng, Jianming Wu, Yanan Wang et al.

Large-scale open-domain dialogue systems such as PLATO-2 have achieved state-of-the-art scores in both English and Chinese. However, little work explores whether such dialogue systems also work well in the Japanese language. In this work, we create a large-scale Japanese dialogue dataset, Dialogue-Graph, which contains 1.656 million dialogue data in a tree structure from News, TV subtitles, and Wikipedia corpus. Then, we train PLATO-2 using Dialogue-Graph to build a large-scale Japanese dialogue system, PLATO-JDS. In addition, to improve the PLATO-JDS in the topic switch issue, we introduce a topic-switch algorithm composed of a topic discriminator to switch to a new topic when user input differs from the previous topic. We evaluate the user experience by using our model with respect to four metrics, namely, coherence, informativeness, engagingness, and humanness. As a result, our proposed PLATO-JDS achieves an average score of 1.500 for the human evaluation with human-bot chat strategy, which is close to the maximum score of 2.000 and suggests the high-quality dialogue generation capability of PLATO-2 in Japanese. Furthermore, our proposed topic-switch algorithm achieves an average score of 1.767 and outperforms PLATO-JDS by 0.267, indicating its effectiveness in improving the user experience of our system.

Yanan Wang, Renxi Wang, Yongxin Wang et al.

Large language models (LLMs) have been extensively studied for tasks like math competitions, complex coding, and scientific reasoning, yet their ability to accurately represent and simulate physical scenarios via code remains underexplored. We propose SimuScene, the first systematic study that trains and evaluates LLMs on simulating physical scenarios across five physics domains and 52 physical concepts. We build an automatic pipeline to collect data, with human verification to ensure quality. The final dataset contains 7,659 physical scenarios with 334 human-verified examples as the test set. We evaluated 10 contemporary LLMs and found that even the strongest model achieves only a 21.5% pass rate, demonstrating the difficulty of the task. Finally, we introduce a reinforcement learning pipeline with visual rewards that uses a vision-language model as a judge to train textual models. Experiments show that training with our data improves physical simulation via code while substantially enhancing general code generation performance.

Yanan Wang, Donghuo Zeng, Shinya Wada et al.

Multimodal transfer learning aims to transform pretrained representations of diverse modalities into a common domain space for effective multimodal fusion. However, conventional systems are typically built on the assumption that all modalities exist, and the lack of modalities always leads to poor inference performance. Furthermore, extracting pretrained embeddings for all modalities is computationally inefficient for inference. In this work, to achieve high efficiency-performance multimodal transfer learning, we propose VideoAdviser, a video knowledge distillation method to transfer multimodal knowledge of video-enhanced prompts from a multimodal fundamental model (teacher) to a specific modal fundamental model (student). With an intuition that the best learning performance comes with professional advisers and smart students, we use a CLIP-based teacher model to provide expressive multimodal knowledge supervision signals to a RoBERTa-based student model via optimizing a step-distillation objective loss -- first step: the teacher distills multimodal knowledge of video-enhanced prompts from classification logits to a regression logit -- second step: the multimodal knowledge is distilled from the regression logit of the teacher to the student. We evaluate our method in two challenging multimodal tasks: video-level sentiment analysis (MOSI and MOSEI datasets) and audio-visual retrieval (VEGAS dataset). The student (requiring only the text modality as input) achieves an MAE score improvement of up to 12.3% for MOSI and MOSEI. Our method further enhances the state-of-the-art method by 3.4% mAP score for VEGAS without additional computations for inference. These results suggest the strengths of our method for achieving high efficiency-performance multimodal transfer learning.

Wanfu Gao, Yanan Wang, Yonghao Li

Multi-label causal feature selection has attracted extensive attention in recent years. However, current methods primarily operate at the label level, treating each label variable as a monolithic entity and overlooking the fine-grained causal mechanisms unique to individual categories. To address this, we propose a Category-level Multi-label Causal Feature selection method named Ca-MCF. Ca-MCF utilizes label category flattening to decompose label variables into specific category nodes, enabling precise modeling of causal structures within the label space. Furthermore, we introduce an explanatory competition-based category-aware recovery mechanism that leverages the proposed Specific Category-Specific Mutual Information (SCSMI) and Distinct Category-Specific Mutual Information (DCSMI) to salvage causal features obscured by label correlations. The method also incorporates structural symmetry checks and cross-dimensional redundancy removal to ensure the robustness and compactness of the identified Markov Blankets. Extensive experiments across seven real-world datasets demonstrate that Ca-MCF significantly outperforms state-of-the-art benchmarks, achieving superior predictive accuracy with reduced feature dimensionality.

Yanan Wang, Yong Ge

Online reviews have played a pivotal role in consumers' decision-making processes. Existing research has highlighted the significant impact of managerial review responses on customer relationship management and firm performance. However, a large portion of online reviews remains unaddressed due to the considerable human labor required to respond to the rapid growth of online reviews. While generative AI has achieved remarkable success in a range of tasks, they are general-purpose models and may not align well with domain-specific human preferences. To tailor these general generative AI models to domain-specific applications, finetuning is commonly employed. Nevertheless, several challenges persist in finetuning with domain-specific data, including hallucinations, difficulty in representing domain-specific human preferences, and over conservatism in offline policy optimization. To address these challenges, we propose a novel preference finetuning method to align an LLM with domain-specific human preferences for generating online review responses. Specifically, we first identify the source of hallucination and propose an effective context augmentation approach to mitigate the LLM hallucination. To represent human preferences, we propose a novel theory-driven preference finetuning approach that automatically constructs human preference pairs in the online review domain. Additionally, we propose a curriculum learning approach to further enhance preference finetuning. To overcome the challenge of over conservatism in existing offline preference finetuning method, we propose a novel density estimation-based support constraint method to relax the conservatism, and we mathematically prove its superior theoretical guarantees. Extensive evaluations substantiate the superiority of our proposed preference finetuning method.

Motonari Kambara, Koki Seno, Tomoya Kaichi et al.

We address language-conditioned robotic manipulation using flow-based trajectory generation, which enables training on human and web videos of object manipulation and requires only minimal embodiment-specific data. This task is challenging, as object trajectory generation from pre-manipulation images and natural language instructions requires appropriate instruction-flow alignment. To tackle this challenge, we propose the flow-based Language Instruction-guided open-Loop ACtion generator (LILAC). This flow-based Vision-Language-Action model (VLA) generates object-centric 2D optical flow from an RGB image and a natural language instruction, and converts the flow into a 6-DoF manipulator trajectory. LILAC incorporates two key components: Semantic Alignment Loss, which strengthens language conditioning to generate instruction-aligned optical flow, and Prompt-Conditioned Cross-Modal Adapter, which aligns learned visual prompts with image and text features to provide rich cues for flow generation. Experimentally, our method outperformed existing approaches in generated flow quality across multiple benchmarks. Furthermore, in physical object manipulation experiments using free-form instructions, LILAC demonstrated a superior task success rate compared to existing methods. The project page is available at https://lilac-75srg.kinsta.page/.

Bing Zhao, Chenfei Wu, Deqing Li et al.

We present Qwen-Image-2.0, an omni-capable image generation foundation model that unifies high-fidelity generation and precise image editing within a single framework. Despite recent progress, existing models still struggle with ultra-long text rendering, multilingual typography, high-resolution photorealism, robust instruction following, and efficient deployment, especially in text-rich and compositionally complex scenarios. Qwen-Image-2.0 addresses these challenges by coupling Qwen3-VL as the condition encoder with a Multimodal Diffusion Transformer for joint condition-target modeling, supported by large-scale data curation and a customized multi-stage training pipeline. This enables strong multimodal understanding while preserving flexible generation and editing capabilities. The model supports instructions of up to 1K tokens for generating text-rich content such as slides, posters, infographics, and comics, while significantly improving multilingual text fidelity and typography. It also enhances photorealistic generation with richer details, more realistic textures, and coherent lighting, and follows complex prompts more reliably across diverse styles. Extensive human evaluations show that Qwen-Image-2.0 substantially outperforms previous Qwen-Image models in both generation and editing, marking a step toward more general, reliable, and practical image generation foundation models.

Heng Guo, Jia Li, Yanan Wang et al.

Construct the first provably secure linear homomorphic ring signature scheme. Ring signatures allow a signer to anonymously sign a message on behalf of a user group (ring) and are widely applied in areas such as identity protection, electronic voting, and privacy enhancement in blockchain. Homomorphic signatures, on the other hand, support verifiable computations on signed data. The integration of anonymity and computability in homomorphic ring signatures holds the potential to create new application scenarios for privacy-preserving distributed systems. It is worth noting that Choi and Kim first introduced the concept of linear homomorphic ring signatures in 2017 and proposed a specific scheme. However, their scheme lacks a complete security proof, leaving its security theoretically unconfirmed. To address this research gap, this paper presents the first provably secure lattice-based linear homomorphic ring signature scheme, designed for scenarios where the ring size is O(log n). This scheme not only combines the anonymity of ring signatures with the malleability of homomorphic signatures but also achieves resistance against quantum attacks.

Yanan Wang, Shengcai Liao, Panwen Hu et al.

Video head swapping aims to replace the entire head of a video subject, including facial identity, head shape, and hairstyle, with that of a reference image, while preserving the target body, background, and motion dynamics. Due to the lack of ground-truth paired swapping data, prior methods typically train on cross-frame pairs of the same person within a video and rely on mask-based inpainting to mitigate identity leakage. Beyond potential boundary artifacts, this paradigm struggles to recover essential cues occluded by the mask, such as facial pose, expressions, and motion dynamics. To address these issues, we prompt a video editing model to synthesize new heads for existing videos as fake swapping inputs, while maintaining frame-synchronized facial poses and expressions. This yields HeadSwapBench, the first cross-identity paired dataset for video head swapping, which supports both training (\TrainNum{} videos) and benchmarking (\TestNum{} videos) with genuine outputs. Leveraging this paired supervision, we propose DirectSwap, a mask-free, direct video head-swapping framework that extends an image U-Net into a video diffusion model with a motion module and conditioning inputs. Furthermore, we introduce the Motion- and Expression-Aware Reconstruction (MEAR) loss, which reweights the diffusion loss per pixel using frame-difference magnitudes and facial-landmark proximity, thereby enhancing cross-frame coherence in motion and expressions. Extensive experiments demonstrate that DirectSwap achieves state-of-the-art visual quality, identity fidelity, and motion and expression consistency across diverse in-the-wild video scenes. We will release the source code and the HeadSwapBench dataset to facilitate future research.

Tengxiao Lv, Ling Luo, Juntao Li et al.

Accurate recognition of biomedical named entities is critical for medical information extraction and knowledge discovery. However, existing methods often struggle with nested entities, entity boundary ambiguity, and cross-lingual generalization. In this paper, we propose a unified Biomedical Named Entity Recognition (BioNER) framework based on Large Language Models (LLMs). We first reformulate BioNER as a text generation task and design a symbolic tagging strategy to jointly handle both flat and nested entities with explicit boundary annotation. To enhance multilingual and multi-task generalization, we perform bilingual joint fine-tuning across multiple Chinese and English datasets. Additionally, we introduce a contrastive learning-based entity selector that filters incorrect or spurious predictions by leveraging boundary-sensitive positive and negative samples. Experimental results on four benchmark datasets and two unseen corpora show that our method achieves state-of-the-art performance and robust zero-shot generalization across languages. The source codes are freely available at https://github.com/dreamer-tx/LLMNER.

Yanan Wang, Zhenghao Fei, Ruichen Li et al.

Recent breakthroughs in large foundation models have enabled the possibility of transferring knowledge pre-trained on vast datasets to domains with limited data availability. Agriculture is one of the domains that lacks sufficient data. This study proposes a framework to train effective, domain-specific, small models from foundation models without manual annotation. Our approach begins with SDM (Segmentation-Description-Matching), a stage that leverages two foundation models: SAM2 (Segment Anything in Images and Videos) for segmentation and OpenCLIP (Open Contrastive Language-Image Pretraining) for zero-shot open-vocabulary classification. In the second stage, a novel knowledge distillation mechanism is utilized to distill compact, edge-deployable models from SDM, enhancing both inference speed and perception accuracy. The complete method, termed SDM-D (Segmentation-Description-Matching-Distilling), demonstrates strong performance across various fruit detection tasks object detection, semantic segmentation, and instance segmentation) without manual annotation. It nearly matches the performance of models trained with abundant labels. Notably, SDM-D outperforms open-set detection methods such as Grounding SAM and YOLO-World on all tested fruit detection datasets. Additionally, we introduce MegaFruits, a comprehensive fruit segmentation dataset encompassing over 25,000 images, and all code and datasets are made publicly available at https://github.com/AgRoboticsResearch/SDM-D.git.

Shuaicong Hu, Yanan Wang, Jian Liu et al.

Considering the variability of amplitude and phase patterns in electrocardiogram (ECG) signals due to cardiac activity and individual differences, existing entropy-based studies have not fully utilized these two patterns and lack integration. To address this gap, this paper proposes a novel fusion entropy metric, morphological ECG entropy (MEE) for the first time, specifically designed for ECG morphology, to comprehensively describe the fusion of amplitude and phase patterns. MEE is computed based on beat-level samples, enabling detailed analysis of each cardiac cycle. Experimental results demonstrate that MEE achieves rapid, accurate, and label-free localization of abnormal ECG arrhythmia regions. Furthermore, MEE provides a method for assessing sample diversity, facilitating compression of imbalanced training sets (via representative sample selection), and outperforms random pruning. Additionally, MEE exhibits the ability to describe areas of poor quality. By discussing, it proves the robustness of MEE value calculation to noise interference and its low computational complexity. Finally, we integrate this method into a clinical interactive interface to provide a more convenient and intuitive user experience. These findings indicate that MEE serves as a valuable clinical descriptor for ECG characterization. The implementation code can be referenced at the following link: https://github.com/fdu-harry/ECG-MEE-metric.

Yanan Wang, Shengcai Liao, Ling Shao

Person re-identification has seen significant advancement in recent years. However, the ability of learned models to generalize to unknown target domains still remains limited. One possible reason for this is the lack of large-scale and diverse source training data, since manually labeling such a dataset is very expensive and privacy sensitive. To address this, we propose to automatically synthesize a large-scale person re-identification dataset following a set-up similar to real surveillance but with virtual environments, and then use the synthesized person images to train a generalizable person re-identification model. Specifically, we design a method to generate a large number of random UV texture maps and use them to create different 3D clothing models. Then, an automatic code is developed to randomly generate various different 3D characters with diverse clothes, races and attributes. Next, we simulate a number of different virtual environments using Unity3D, with customized camera networks similar to real surveillance systems, and import multiple 3D characters at the same time, with various movements and interactions along different paths through the camera networks. As a result, we obtain a virtual dataset, called RandPerson, with 1,801,816 person images of 8,000 identities. By training person re-identification models on these synthesized person images, we demonstrate, for the first time, that models trained on virtual data can generalize well to unseen target images, surpassing the models trained on various real-world datasets, including CUHK03, Market-1501, DukeMTMC-reID, and almost MSMT17. The RandPerson dataset is available at https://github.com/VideoObjectSearch/RandPerson.

Yanan Wang, Shuichiro Haruta, Donghuo Zeng et al.

Video question answering (VideoQA) is a task to predict the correct answer to questions posed about a given video. The system must comprehend spatial and temporal relationships among objects extracted from videos to perform causal and temporal reasoning. While prior works have focused on modeling individual object movements using transformer-based methods, they falter when capturing complex scenarios involving multiple objects (e.g., "a boy is throwing a ball in a hoop"). We propose a contrastive language event graph representation learning method called CLanG to address this limitation. Aiming to capture event representations associated with multiple objects, our method employs a multi-layer GNN-cluster module for adversarial graph representation learning, enabling contrastive learning between the question text and its relevant multi-object event graph. Our method outperforms a strong baseline, achieving up to 2.2% higher accuracy on two challenging VideoQA datasets, NExT-QA and TGIF-QA-R. In particular, it is 2.8% better than baselines in handling causal and temporal questions, highlighting its strength in reasoning multiple object-based events.

Mengyuan Tian, Qiyan Zhao, Yanan Wang et al.

In recent years, pre-trained visual-linguistic models have demonstrated tremendous potential, becoming a crucial foundational framework for numerous downstream tasks. However, the information density between text and images is not uniformly distributed. Existing methods often overlook the inherent and dynamic differences in information density and semantic scope between text tags and image blocks. These common uniform alignment strategies result in coarse-grained cross-modal interactions and loss of fine semantic details. Moreover, pursuing finer alignment typically requires substantial computational overhead, limiting practical model deployment. To address this challenge, this paper proposes a novel framework for dynamic cross-modal alignment with continuous detail introduction. First, we design a dynamically adaptive cross-modal matching mechanism that uses a learnable matching function to dynamically assign varying numbers and sizes of image tags to text tags of the same size but different information density, enabling more precise attention interaction. Second, we develop a continuous detail introduction module to progressively incorporate high-resolution visual feature enhancement into the alignment process. Extensive experiments across multiple benchmarks demonstrate significant improvements in the accuracy of various downstream tasks while reducing computational overhead.

Yanan Wang, Shuichiro Haruta, Donghuo Zeng et al.

Capturing complex hierarchical human activities, from atomic actions (e.g., picking up one present, moving to the sofa, unwrapping the present) to contextual events (e.g., celebrating Christmas) is crucial for achieving high-performance video question answering (VideoQA). Recent works have expanded multimodal models (e.g., CLIP, LLaVA) to process continuous video sequences, enhancing the model's temporal reasoning capabilities. However, these approaches often fail to capture contextual events that can be decomposed into multiple atomic actions non-continuously distributed over relatively long-term sequences. In this paper, to leverage the spatial visual context representation capability of the CLIP model for obtaining non-continuous visual representations in terms of contextual events in videos, we convert long-term video sequences into a spatial image domain and finetune the multimodal model LLaVA for the VideoQA task. Our approach achieves competitive performance on the STAR task, in particular, with a 78.4% accuracy score, exceeding the current state-of-the-art score by 2.8 points on the NExTQA task.

Yanan Wang, Linjie Ren, Zihao Li et al.

While video-to-audio generation has achieved remarkable progress in semantic and temporal alignment, most existing studies focus solely on these aspects, paying limited attention to the spatial perception and immersive quality of the synthesized audio. This limitation stems largely from current models' reliance on mono audio datasets, which lack the binaural spatial information needed to learn visual-to-spatial audio mappings. To address this gap, we introduce two key contributions: we construct BinauralVGGSound, the first large-scale video-binaural audio dataset designed to support spatially aware video-to-audio generation; and we propose a end-to-end spatial audio generation framework guided by visual cues, which explicitly models spatial features. Our framework incorporates a visual-guided audio spatialization module that ensures the generated audio exhibits realistic spatial attributes and layered spatial depth while maintaining semantic and temporal alignment. Experiments show that our approach substantially outperforms state-of-the-art models in spatial fidelity and delivers a more immersive auditory experience, without sacrificing temporal or semantic consistency. All datasets, code, and model checkpoints will be publicly released to facilitate future research.

Satoshi Hashimoto, Hitoshi Nishimura, Yanan Wang et al.

Deploying video anomaly detection in practice is hampered by the scarcity and collection cost of real abnormal footage. We address this by training without any real abnormal videos while evaluating under the standard weakly supervised split, and we introduce PA-VAD, a generation-driven approach that learns a detector from synthesized pseudo-abnormal videos paired with real normal videos, using only a small set of real normal images to drive synthesis. For synthesis, we select class-relevant initial images with CLIP and refine textual prompts with a vision-language model to improve fidelity and scene consistency before invoking a video diffusion model. For training, we mitigate excessive spatiotemporal magnitude in synthesized anomalies by an domain-aligned regularized module that combines domain alignment and memory usage-aware updates. Extensive experiments show that our approach reaches 98.2% on ShanghaiTech and 82.5% on UCF-Crime, surpassing the strongest real-abnormal method on ShanghaiTech by +0.6% and outperforming the UVAD state-of-the-art on UCF-Crime by +1.9%. The results demonstrate that high-accuracy anomaly detection can be obtained without collecting real anomalies, providing a practical path toward scalable deployment.

Donghuo Zeng, Yanan Wang, Kazushi Ikeda et al.

Metric learning minimizes the gap between similar (positive) pairs of data points and increases the separation of dissimilar (negative) pairs, aiming at capturing the underlying data structure and enhancing the performance of tasks like audio-visual cross-modal retrieval (AV-CMR). Recent works employ sampling methods to select impactful data points from the embedding space during training. However, the model training fails to fully explore the space due to the scarcity of training data points, resulting in an incomplete representation of the overall positive and negative distributions. In this paper, we propose an innovative Anchor-aware Deep Metric Learning (AADML) method to address this challenge by uncovering the underlying correlations among existing data points, which enhances the quality of the shared embedding space. Specifically, our method establishes a correlation graph-based manifold structure by considering the dependencies between each sample as the anchor and its semantically similar samples. Through dynamic weighting of the correlations within this underlying manifold structure using an attention-driven mechanism, Anchor Awareness (AA) scores are obtained for each anchor. These AA scores serve as data proxies to compute relative distances in metric learning approaches. Extensive experiments conducted on two audio-visual benchmark datasets demonstrate the effectiveness of our proposed AADML method, significantly surpassing state-of-the-art models. Furthermore, we investigate the integration of AA proxies with various metric learning methods, further highlighting the efficacy of our approach.

Paridhi Maheshwari, Hongyu Ren, Yanan Wang et al.

Many real-world systems exhibit temporal, dynamic behaviors, which are captured as time series of complex agent interactions. To perform temporal reasoning, current methods primarily encode temporal dynamics through simple sequence-based models. However, in general these models fail to efficiently capture the full spectrum of rich dynamics in the input, since the dynamics is not uniformly distributed. In particular, relevant information might be harder to extract and computing power is wasted for processing all individual timesteps, even if they contain no significant changes or no new information. Here we propose TimeGraphs, a novel approach that characterizes dynamic interactions as a hierarchical temporal graph, diverging from traditional sequential representations. Our approach models the interactions using a compact graph-based representation, enabling adaptive reasoning across diverse time scales. Adopting a self-supervised method, TimeGraphs constructs a multi-level event hierarchy from a temporal input, which is then used to efficiently reason about the unevenly distributed dynamics. This construction process is scalable and incremental to accommodate streaming data. We evaluate TimeGraphs on multiple datasets with complex, dynamic agent interactions, including a football simulator, the Resistance game, and the MOMA human activity dataset. The results demonstrate both robustness and efficiency of TimeGraphs on a range of temporal reasoning tasks. Our approach obtains state-of-the-art performance and leads to a performance increase of up to 12.2% on event prediction and recognition tasks over current approaches. Our experiments further demonstrate a wide array of capabilities including zero-shot generalization, robustness in case of data sparsity, and adaptability to streaming data flow.

Peng Xu, Shengwu Xiong, Jiajun Zhang et al.

This paper reviews the MARS2 2025 Challenge on Multimodal Reasoning. We aim to bring together different approaches in multimodal machine learning and LLMs via a large benchmark. We hope it better allows researchers to follow the state-of-the-art in this very dynamic area. Meanwhile, a growing number of testbeds have boosted the evolution of general-purpose large language models. Thus, this year's MARS2 focuses on real-world and specialized scenarios to broaden the multimodal reasoning applications of MLLMs. Our organizing team released two tailored datasets Lens and AdsQA as test sets, which support general reasoning in 12 daily scenarios and domain-specific reasoning in advertisement videos, respectively. We evaluated 40+ baselines that include both generalist MLLMs and task-specific models, and opened up three competition tracks, i.e., Visual Grounding in Real-world Scenarios (VG-RS), Visual Question Answering with Spatial Awareness (VQA-SA), and Visual Reasoning in Creative Advertisement Videos (VR-Ads). Finally, 76 teams from the renowned academic and industrial institutions have registered and 40+ valid submissions (out of 1200+) have been included in our ranking lists. Our datasets, code sets (40+ baselines and 15+ participants' methods), and rankings are publicly available on the MARS2 workshop website and our GitHub organization page https://github.com/mars2workshop/, where our updates and announcements of upcoming events will be continuously provided.

Yanan Wang, Julio Vizcarra, Zhi Li et al.

Despite recent progress in video large language models (VideoLLMs), a key open challenge remains: how to equip models with chain-of-thought (CoT) reasoning abilities grounded in fine-grained object-level video understanding. Existing instruction-tuned models, such as the Qwen and LLaVA series, are trained on high-level video-text pairs, often lacking structured annotations necessary for compositional, step-by-step reasoning. We propose CoTasks: Chain-of-Thought based Video Instruction Tuning Tasks, a new framework that decomposes complex video questions of existing datasets (e.g., NeXT-QA, STAR) into four entity-level foundational tasks: frame localization, entity tracking, spatial and temporal relation extraction. By embedding these intermediate CoT-style reasoning steps into the input, CoTasks enables models to explicitly perform object-centric spatiotemporal reasoning. Experiments on the NeXT-QA benchmark show that CoTasks significantly enhance inference performance: LLaVA-video-7B improves by +3.3 points in average GPT-4 evaluation score, and Qwen2.5-VL-3B gains +17.4, with large boosts in causal (+14.6), temporal (+10.9), and descriptive (+48.1) subcategories. These results demonstrate the effectiveness of CoTasks as a structured CoT-style supervision framework for improving compositional video reasoning.

Shuichiro Haruta, Kazunori Matsumoto, Zhi Li et al.

In this paper, we propose a rotation-constrained compensation method to address the errors introduced by structured pruning of large language models (LLMs). LLMs are trained on massive datasets and accumulate rich semantic knowledge in their representation space. In contrast, pruning is typically carried out with only a small amount of calibration data, which makes output mismatches unavoidable. Although direct least-squares fitting can reduce such errors, it tends to overfit to the limited calibration set, destructively modifying pretrained weights. To overcome this difficulty, we update the pruned parameters under a rotation constraint. This constrained update preserves the geometry of output representations (i.e., norms and inner products) and simultaneously re-aligns the pruned subspace with the original outputs. Furthermore, in rotation-constrained compensation, removing components that strongly contribute to the principal directions of the output makes error recovery difficult. Since input dimensions with large variance strongly affect these principal directions, we design a variance-aware importance score that ensures such dimensions are preferentially kept in the pruned model. By combining this scoring rule with rotation-constrained updates, the proposed method effectively compensates errors while retaining the components likely to be more important in a geometry-preserving manner. In the experiments, we apply the proposed method to LLaMA-7B and evaluate it on WikiText-2 and multiple language understanding benchmarks. The results demonstrate consistently better perplexity and task accuracy compared with existing baselines.

Rui Wang, Xing Liu, Yanan Wang et al.

The recently released artificial intelligence conversational agent, ChatGPT, has gained significant attention in academia and real life. A multitude of early ChatGPT users eagerly explore its capabilities and share their opinions on it via social media. Both user queries and social media posts express public concerns regarding this advanced dialogue system. To mine public concerns about ChatGPT, a novel Self-Supervised neural Topic Model (SSTM), which formalizes topic modeling as a representation learning procedure, is proposed in this paper. Extensive experiments have been conducted on Twitter posts about ChatGPT and queries asked by ChatGPT users. And experimental results demonstrate that the proposed approach could extract higher quality public concerns with improved interpretability and diversity, surpassing the performance of state-of-the-art approaches.

Dadong Miao, Yanan Wang, Guoyu Tang et al.

Recent years have seen a significant amount of interests in Sequential Recommendation (SR), which aims to understand and model the sequential user behaviors and the interactions between users and items over time. Surprisingly, despite the huge success Sequential Recommendation has achieved, there is little study on Sequential Search (SS), a twin learning task that takes into account a user's current and past search queries, in addition to behavior on historical query sessions. The SS learning task is even more important than the counterpart SR task for most of E-commence companies due to its much larger online serving demands as well as traffic volume. To this end, we propose a highly scalable hybrid learning model that consists of an RNN learning framework leveraging all features in short-term user-item interactions, and an attention model utilizing selected item-only features from long-term interactions. As a novel optimization step, we fit multiple short user sequences in a single RNN pass within a training batch, by solving a greedy knapsack problem on the fly. Moreover, we explore the use of off-policy reinforcement learning in multi-session personalized search ranking. Specifically, we design a pairwise Deep Deterministic Policy Gradient model that efficiently captures users' long term reward in terms of pairwise classification error. Extensive ablation experiments demonstrate significant improvement each component brings to its state-of-the-art baseline, on a variety of offline and online metrics.

Haowei He, Jingzhao Zhang, Yanan Wang et al.

Electric vehicles (EVs) play an important role in reducing carbon emissions. As EV adoption accelerates, safety issues caused by EV batteries have become an important research topic. In order to benchmark and develop data-driven methods for this task, we introduce a large and comprehensive dataset of EV batteries. Our dataset includes charging records collected from hundreds of EVs from three manufacturers over several years. Our dataset is the first large-scale public dataset on real-world battery data, as existing data either include only several vehicles or is collected in the lab environment. Meanwhile, our dataset features two types of labels, corresponding to two key tasks - battery health estimation and battery capacity estimation. In addition to demonstrating how existing deep learning algorithms can be applied to this task, we further develop an algorithm that exploits the data structure of battery systems. Our algorithm achieves better results and shows that a customized method can improve model performances. We hope that this public dataset provides valuable resources for researchers, policymakers, and industry professionals to better understand the dynamics of EV battery aging and support the transition toward a sustainable transportation system.

Yanan Wang, Yong Ge, Li Li et al.

Reinforcement learning (RL) has shown great promise in optimizing long-term user interest in recommender systems. However, existing RL-based recommendation methods need a large number of interactions for each user to learn a robust recommendation policy. The challenge becomes more critical when recommending to new users who have a limited number of interactions. To that end, in this paper, we address the cold-start challenge in the RL-based recommender systems by proposing a meta-level model-based reinforcement learning approach for fast user adaptation. In our approach, we learn to infer each user's preference with a user context variable that enables recommendation systems to better adapt to new users with few interactions. To improve adaptation efficiency, we learn to recover the user policy and reward from only a few interactions via an inverse reinforcement learning method to assist a meta-level recommendation agent. Moreover, we model the interaction relationship between the user model and recommendation agent from an information-theoretic perspective. Empirical results show the effectiveness of the proposed method when adapting to new users with only a single interaction sequence. We further provide a theoretical analysis of the recommendation performance bound.

Xiaoyu Zeng, Yanan Wang, Tai-Yin Chiu et al.

The task of answering questions about images has garnered attention as a practical service for assisting populations with visual impairments as well as a visual Turing test for the artificial intelligence community. Our first aim is to identify the common vision skills needed for both scenarios. To do so, we analyze the need for four vision skills---object recognition, text recognition, color recognition, and counting---on over 27,000 visual questions from two datasets representing both scenarios. We next quantify the difficulty of these skills for both humans and computers on both datasets. Finally, we propose a novel task of predicting what vision skills are needed to answer a question about an image. Our results reveal (mis)matches between aims of real users of such services and the focus of the AI community. We conclude with a discussion about future directions for addressing the visual question answering task.

Yanan Wang, Tong Xu, Xin Niu et al.

The development of intelligent traffic light control systems is essential for smart transportation management. While some efforts have been made to optimize the use of individual traffic lights in an isolated way, related studies have largely ignored the fact that the use of multi-intersection traffic lights is spatially influenced and there is a temporal dependency of historical traffic status for current traffic light control. To that end, in this paper, we propose a novel SpatioTemporal Multi-Agent Reinforcement Learning (STMARL) framework for effectively capturing the spatio-temporal dependency of multiple related traffic lights and control these traffic lights in a coordinating way. Specifically, we first construct the traffic light adjacency graph based on the spatial structure among traffic lights. Then, historical traffic records will be integrated with current traffic status via Recurrent Neural Network structure. Moreover, based on the temporally-dependent traffic information, we design a Graph Neural Network based model to represent relationships among multiple traffic lights, and the decision for each traffic light will be made in a distributed way by the deep Q-learning method. Finally, the experimental results on both synthetic and real-world data have demonstrated the effectiveness of our STMARL framework, which also provides an insightful understanding of the influence mechanism among multi-intersection traffic lights.