Dongmei Jiang

Tao Jiang, Xinmeng Yu, Chenhao Yi et al.

Evolutionary model merging provides a powerful framework for the automated, training-free composition of LLMs through parameter-space search. However, existing methods predominantly rely on stochastic, hand-crafted operators that overlook the underlying performance landscape of the coefficient space. We propose Evolutionary Generative Merging (EvoGM), a framework that transcends manual heuristics by employing learnable generative modeling to optimize merging coefficients. Specifically, EvoGM features a dual-generator architecture with cycle-consistent learning to adaptively sample and refine promising merging candidates. By constructing winner-loser pairs from historical search trajectories, our framework effectively captures high-performance parameter distributions and maximizes data efficiency. This generative process is seamlessly integrated into a multi-round evolutionary pipeline, where elite merged models iteratively serve as new expert foundations. Extensive experiments across diverse benchmarks demonstrate that EvoGM significantly outperforms state-of-the-art baselines, exhibiting robust performance on both seen and unseen tasks. Code and data are available at https://github.com/JiangTao97/evogm.

Hao Wang, Pengzhen Ren, Zequn Jie et al.

Open-vocabulary detection is a challenging task due to the requirement of detecting objects based on class names, including those not encountered during training. Existing methods have shown strong zero-shot detection capabilities through pre-training and pseudo-labeling on diverse large-scale datasets. However, these approaches encounter two main challenges: (i) how to effectively eliminate data noise from pseudo-labeling, and (ii) how to efficiently leverage the language-aware capability for region-level cross-modality fusion and alignment. To address these challenges, we propose a novel unified open-vocabulary detection method called OV-DINO, which is pre-trained on diverse large-scale datasets with language-aware selective fusion in a unified framework. Specifically, we introduce a Unified Data Integration (UniDI) pipeline to enable end-to-end training and eliminate noise from pseudo-label generation by unifying different data sources into detection-centric data format. In addition, we propose a Language-Aware Selective Fusion (LASF) module to enhance the cross-modality alignment through a language-aware query selection and fusion process. We evaluate the performance of the proposed OV-DINO on popular open-vocabulary detection benchmarks, achieving state-of-the-art results with an AP of 50.6% on the COCO benchmark and 40.1% on the LVIS benchmark in a zero-shot manner, demonstrating its strong generalization ability. Furthermore, the fine-tuned OV-DINO on COCO achieves 58.4% AP, outperforming many existing methods with the same backbone. The code for OV-DINO is available at https://github.com/wanghao9610/OV-DINO.

Guiping Cao, Shengda Luo, Wenjian Huang et al.

Token interaction operation is one of the core modules in MLP-based models to exchange and aggregate information between different spatial locations. However, the power of token interaction on the spatial dimension is highly dependent on the spatial resolution of the feature maps, which limits the model's expressive ability, especially in deep layers where the feature are down-sampled to a small spatial size. To address this issue, we present a novel method called \textbf{Strip-MLP} to enrich the token interaction power in three ways. Firstly, we introduce a new MLP paradigm called Strip MLP layer that allows the token to interact with other tokens in a cross-strip manner, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). Secondly, a \textbf{C}ascade \textbf{G}roup \textbf{S}trip \textbf{M}ixing \textbf{M}odule (CGSMM) is proposed to overcome the performance degradation caused by small spatial feature size. The module allows tokens to interact more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. Finally, based on the Strip MLP layer, we propose a novel \textbf{L}ocal \textbf{S}trip \textbf{M}ixing \textbf{M}odule (LSMM) to boost the token interaction power in the local region. Extensive experiments demonstrate that Strip-MLP significantly improves the performance of MLP-based models on small datasets and obtains comparable or even better results on ImageNet. In particular, Strip-MLP models achieve higher average Top-1 accuracy than existing MLP-based models by +2.44\% on Caltech-101 and +2.16\% on CIFAR-100. The source codes will be available at~\href{https://github.com/Med-Process/Strip_MLP{https://github.com/Med-Process/Strip\_MLP}.

Qiaohui Chu, Haoyu Zhang, Yisen Feng et al.

We propose JFAA, a JEPA-based Future Action Anticipation method for the EPIC-KITCHENS-100 (EK-100) Action Anticipation task. Inspired by the representation learning and future prediction ability of V-JEPA 2.1, JFAA uses a frozen encoder and predictor to extract observed context features and near-future latent tokens. A lightweight attentive probe is then trained to predict verb, noun, and action logits with separate task queries. To improve robustness, we further build a field-aware ensemble over selected epoch-level predictions, allowing each output field to benefit from its most reliable candidates. Experimental results on the official challenge server show that JFAA achieves first place in the EgoVis 2026 EK-100 Action Anticipation Challenge. Our code will be released at https://github.com/CorrineQiu/JFAA.

Qiaohui Chu, Haoyu Zhang, Yisen Feng et al.

We propose VISTA, a V-JEPA Integrated StillFast Temporal Anticipator for the Ego4D Short-Term Object Interaction Anticipation (STA) Challenge at EgoVis 2026. Given an egocentric video timestamp, the task requires anticipating the next human-object interaction, including the future active object's bounding box, noun category, verb category, time-to-contact, and confidence score. VISTA follows a StillFast-style design that combines object-centric spatial detection with short-horizon temporal context. Specifically, a COCO-pretrained Faster R-CNN ResNet-50 FPN detector generates object proposals from the last observed high-resolution frame, while a frozen V-JEPA 2.1 temporal branch extracts clip-level egocentric context from the observed video. The temporal representation is injected into the detection pathway through feature modulation and ROI-level context fusion. The fused proposal features are then passed to multi-head STA predictors for box refinement, noun classification, verb classification, time-to-contact regression, and interaction confidence estimation. For the final submission, we further ensemble complementary predictions to improve robustness. Experimental results on the official challenge server show that VISTA achieves first place in the EgoVis 2026 Ego4D STA Challenge. Our code will be released at https://github.com/CorrineQiu/VISTA.

Zaijing Li, Yuquan Xie, Rui Shao et al.

Building a general-purpose agent is a long-standing vision in the field of artificial intelligence. Existing agents have made remarkable progress in many domains, yet they still struggle to complete long-horizon tasks in an open world. We attribute this to the lack of necessary world knowledge and multimodal experience that can guide agents through a variety of long-horizon tasks. In this paper, we propose a Hybrid Multimodal Memory module to address the above challenges. It 1) transforms knowledge into Hierarchical Directed Knowledge Graph that allows agents to explicitly represent and learn world knowledge, and 2) summarises historical information into Abstracted Multimodal Experience Pool that provide agents with rich references for in-context learning. On top of the Hybrid Multimodal Memory module, a multimodal agent, Optimus-1, is constructed with dedicated Knowledge-guided Planner and Experience-Driven Reflector, contributing to a better planning and reflection in the face of long-horizon tasks in Minecraft. Extensive experimental results show that Optimus-1 significantly outperforms all existing agents on challenging long-horizon task benchmarks, and exhibits near human-level performance on many tasks. In addition, we introduce various Multimodal Large Language Models (MLLMs) as the backbone of Optimus-1. Experimental results show that Optimus-1 exhibits strong generalization with the help of the Hybrid Multimodal Memory module, outperforming the GPT-4V baseline on many tasks.

Yi Zhang, Jitao Sang, Junyang Wang et al.

Machine learning models often learn to make predictions that rely on sensitive social attributes like gender and race, which poses significant fairness risks, especially in societal applications, such as hiring, banking, and criminal justice. Existing work tackles this issue by minimizing the employed information about social attributes in models for debiasing. However, the high correlation between target task and these social attributes makes learning on the target task incompatible with debiasing. Given that model bias arises due to the learning of bias features (\emph{i.e}., gender) that help target task optimization, we explore the following research question: \emph{Can we leverage shortcut features to replace the role of bias feature in target task optimization for debiasing?} To this end, we propose \emph{Shortcut Debiasing}, to first transfer the target task's learning of bias attributes from bias features to shortcut features, and then employ causal intervention to eliminate shortcut features during inference. The key idea of \emph{Shortcut Debiasing} is to design controllable shortcut features to on one hand replace bias features in contributing to the target task during the training stage, and on the other hand be easily removed by intervention during the inference stage. This guarantees the learning of the target task does not hinder the elimination of bias features. We apply \emph{Shortcut Debiasing} to several benchmark datasets, and achieve significant improvements over the state-of-the-art debiasing methods in both accuracy and fairness.

Zheng Chong, Xiao Dong, Haoxiang Li et al.

Virtual try-on methods based on diffusion models achieve realistic effects but often require additional encoding modules, a large number of training parameters, and complex preprocessing, which increases the burden on training and inference. In this work, we re-evaluate the necessity of additional modules and analyze how to improve training efficiency and reduce redundant steps in the inference process. Based on these insights, we propose CatVTON, a simple and efficient virtual try-on diffusion model that transfers in-shop or worn garments of arbitrary categories to target individuals by concatenating them along spatial dimensions as inputs of the diffusion model. The efficiency of CatVTON is reflected in three aspects: (1) Lightweight network. CatVTON consists only of a VAE and a simplified denoising UNet, removing redundant image and text encoders as well as cross-attentions, and includes just 899.06M parameters. (2) Parameter-efficient training. Through experimental analysis, we identify self-attention modules as crucial for adapting pre-trained diffusion models to the virtual try-on task, enabling high-quality results with only 49.57M training parameters. (3) Simplified inference. CatVTON eliminates unnecessary preprocessing, such as pose estimation, human parsing, and captioning, requiring only a person image and garment reference to guide the virtual try-on process, reducing over 49% memory usage compared to other diffusion-based methods. Extensive experiments demonstrate that CatVTON achieves superior qualitative and quantitative results compared to baseline methods and demonstrates strong generalization performance in in-the-wild scenarios, despite being trained solely on public datasets with 73K samples.

Fan Cui, Liyong Guo, Lang He et al.

Electroencephalography (EEG) plays a vital role in detecting how brain responses to different stimulus. In this paper, we propose a novel Shallow-Deep Attention-based Network (SDANet) to classify the correct auditory stimulus evoking the EEG signal. It adopts the Attention-based Correlation Module (ACM) to discover the connection between auditory speech and EEG from global aspect, and the Shallow-Deep Similarity Classification Module (SDSCM) to decide the classification result via the embeddings learned from the shallow and deep layers. Moreover, various training strategies and data augmentation are used to boost the model robustness. Experiments are conducted on the dataset provided by Auditory EEG challenge (ICASSP Signal Processing Grand Challenge 2023). Results show that the proposed model has a significant gain over the baseline on the match-mismatch track.

Feng Lu, Xiangyuan Lan, Lijun Zhang et al.

Over the past decade, most methods in visual place recognition (VPR) have used neural networks to produce feature representations. These networks typically produce a global representation of a place image using only this image itself and neglect the cross-image variations (e.g. viewpoint and illumination), which limits their robustness in challenging scenes. In this paper, we propose a robust global representation method with cross-image correlation awareness for VPR, named CricaVPR. Our method uses the attention mechanism to correlate multiple images within a batch. These images can be taken in the same place with different conditions or viewpoints, or even captured from different places. Therefore, our method can utilize the cross-image variations as a cue to guide the representation learning, which ensures more robust features are produced. To further facilitate the robustness, we propose a multi-scale convolution-enhanced adaptation method to adapt pre-trained visual foundation models to the VPR task, which introduces the multi-scale local information to further enhance the cross-image correlation-aware representation. Experimental results show that our method outperforms state-of-the-art methods by a large margin with significantly less training time. The code is released at https://github.com/Lu-Feng/CricaVPR.

Xing Lan, Jian Xue, Ji Qi et al.

Facial expression recognition (FER) is a critical task in multimedia with significant implications across various domains. However, analyzing the causes of facial expressions is essential for accurately recognizing them. Current approaches, such as those based on facial action units (AUs), typically provide AU names and intensities but lack insight into the interactions and relationships between AUs and the overall expression. In this paper, we propose a novel method called ExpLLM, which leverages large language models to generate an accurate chain of thought (CoT) for facial expression recognition. Specifically, we have designed the CoT mechanism from three key perspectives: key observations, overall emotional interpretation, and conclusion. The key observations describe the AU's name, intensity, and associated emotions. The overall emotional interpretation provides an analysis based on multiple AUs and their interactions, identifying the dominant emotions and their relationships. Finally, the conclusion presents the final expression label derived from the preceding analysis. Furthermore, we also introduce the Exp-CoT Engine, designed to construct this expression CoT and generate instruction-description data for training our ExpLLM. Extensive experiments on the RAF-DB and AffectNet datasets demonstrate that ExpLLM outperforms current state-of-the-art FER methods. ExpLLM also surpasses the latest GPT-4o in expression CoT generation, particularly in recognizing micro-expressions where GPT-4o frequently fails.

Feng Lu, Shuting Dong, Lijun Zhang et al.

Visual place recognition (VPR) is a fundamental task for many applications such as robot localization and augmented reality. Recently, the hierarchical VPR methods have received considerable attention due to the trade-off between accuracy and efficiency. They usually first use global features to retrieve the candidate images, then verify the spatial consistency of matched local features for re-ranking. However, the latter typically relies on the RANSAC algorithm for fitting homography, which is time-consuming and non-differentiable. This makes existing methods compromise to train the network only in global feature extraction. Here, we propose a transformer-based deep homography estimation (DHE) network that takes the dense feature map extracted by a backbone network as input and fits homography for fast and learnable geometric verification. Moreover, we design a re-projection error of inliers loss to train the DHE network without additional homography labels, which can also be jointly trained with the backbone network to help it extract the features that are more suitable for local matching. Extensive experiments on benchmark datasets show that our method can outperform several state-of-the-art methods. And it is more than one order of magnitude faster than the mainstream hierarchical VPR methods using RANSAC. The code is released at https://github.com/Lu-Feng/DHE-VPR.

Wenyun Li, Zheng Zhang, Dongmei Jiang et al.

Vision Transformer (ViT) models have achieved remarkable performance across various vision tasks, with scalability being a key advantage when applied to large datasets. This scalability enables ViT models to exhibit strong generalization capabilities. However, as the number of parameters increases, the robustness of ViT models to adversarial examples does not scale proportionally. Adversarial training (AT), one of the most effective methods for enhancing robustness, typically requires fine-tuning the entire model, leading to prohibitively high computational costs, especially for large ViT architectures. In this paper, we aim to robustly fine-tune only a small subset of parameters to achieve robustness comparable to standard AT. To accomplish this, we introduce Criticality-Aware Adversarial Training (CAAT), a novel method that adaptively allocates resources to the most robustness-critical parameters, fine-tuning only selected modules. Specifically, CAAT efficiently identifies parameters that contribute most to adversarial robustness. It then leverages parameter-efficient fine-tuning (PEFT) to robustly adjust weight matrices where the number of critical parameters exceeds a predefined threshold. CAAT exhibits favorable generalization when scaled to larger vision transformer architectures, potentially paving the way for adversarial training at scale, e.g, compared with plain adversarial training, CAAT incurs only a 4.3% decrease in adversarial robustness while tuning approximately 6% of its parameters. Extensive experiments on three widely used adversarial learning datasets demonstrate that CAAT outperforms state-of-the-art lightweight AT methods with fewer trainable parameters.

Yan Li, Yifei Xing, Xiangyuan Lan et al.

In the era of large-scale pre-trained models, effectively adapting general knowledge to specific affective computing tasks remains a challenge, particularly regarding computational efficiency and multimodal heterogeneity. While Transformer-based methods have excelled at modeling inter-modal dependencies, their quadratic computational complexity limits their use with long-sequence data. Mamba-based models have emerged as a computationally efficient alternative; however, their inherent sequential scanning mechanism struggles to capture the global, non-sequential relationships that are crucial for effective cross-modal alignment. To address these limitations, we propose \textbf{AlignMamba-2}, an effective and efficient framework for multimodal fusion and sentiment analysis. Our approach introduces a dual alignment strategy that regularizes the model using both Optimal Transport distance and Maximum Mean Discrepancy, promoting geometric and statistical consistency between modalities without incurring any inference-time overhead. More importantly, we design a Modality-Aware Mamba layer, which employs a Mixture-of-Experts architecture with modality-specific and modality-shared experts to explicitly handle data heterogeneity during the fusion process. Extensive experiments on four challenging benchmarks, including dynamic time-series (on the CMU-MOSI and CMU-MOSEI datasets) and static image-related tasks (on the NYU-Depth V2 and MVSA-Single datasets), demonstrate that AlignMamba-2 establishes a new state-of-the-art in both effectiveness and efficiency across diverse pattern recognition tasks, ranging from dynamic time-series analysis to static image-text classification.

Guiping Cao, Tao Wang, Wenjian Huang et al.

Open-Ended object Detection (OED) is a novel and challenging task that detects objects and generates their category names in a free-form manner, without requiring additional vocabularies during inference. However, the existing OED models, such as GenerateU, require large-scale datasets for training, suffer from slow convergence, and exhibit limited performance. To address these issues, we present a novel and efficient Open-Det framework, consisting of four collaborative parts. Specifically, Open-Det accelerates model training in both the bounding box and object name generation process by reconstructing the Object Detector and the Object Name Generator. To bridge the semantic gap between Vision and Language modalities, we propose a Vision-Language Aligner with V-to-L and L-to-V alignment mechanisms, incorporating with the Prompts Distiller to transfer knowledge from the VLM into VL-prompts, enabling accurate object name generation for the LLM. In addition, we design a Masked Alignment Loss to eliminate contradictory supervision and introduce a Joint Loss to enhance classification, resulting in more efficient training. Compared to GenerateU, Open-Det, using only 1.5% of the training data (0.077M vs. 5.077M), 20.8% of the training epochs (31 vs. 149), and fewer GPU resources (4 V100 vs. 16 A100), achieves even higher performance (+1.0% in APr). The source codes are available at: https://github.com/Med-Process/Open-Det.

Mingchen Song, Xiang Deng, Jie Wei et al.

Recent advances in unimanual manipulation policies have achieved remarkable success across diverse robotic tasks through abundant training data and well-established model architectures. However, extending these capabilities to bimanual manipulation remains challenging due to the lack of bimanual demonstration data and the complexity of coordinating dual-arm actions. Existing approaches either rely on extensive bimanual datasets or fail to effectively leverage pre-trained unimanual policies. To address this limitation, we propose \textbf{EnergyAction}, a novel framework that compositionally transfers unimanual manipulation policies to bimanual tasks through the Energy-Based Models (EBMs). Specifically, our method incorporates three key innovations. First, we model individual unimanual policies as EBMs and leverage their compositional properties to compose left and right arm actions, enabling the fusion of unimanual policies into a bimanual policy. Second, we introduce an energy-based temporal-spatial coordination mechanism through energy constraints, ensuring the generated bimanual actions are both temporal coherence and spatial feasibility. Third, we propose two different energy-aware denoising strategies that dynamically adapt denoising steps based on action quality assessment. These strategies ensure the generation of high-quality actions while maintaining superior computational efficiency compared to fixed-step denoising approaches. Experimental results demonstrate that EnergyAction effectively transfers unimanual knowledge to bimanual tasks, achieving superior performance on both simulated and real-world tasks with minimal bimanual data.

Guiping Cao, Xiangyuan Lan, Wenjian Huang et al.

Popular transformer detectors have achieved promising performance through query-based learning using attention mechanisms. However, the roles of existing decoder query types (e.g., content query and positional query) are still underexplored. These queries are generally predefined with a fixed number (fixed-query), which limits their flexibility. We find that the learning of these fixed-query is impaired by Recurrent Opposing inTeractions (ROT) between two attention operations: Self-Attention (query-to-query) and Cross-Attention (query-to-encoder), thereby degrading decoder efficiency. Furthermore, "query ambiguity" arises when shared-weight decoder layers are processed with both one-to-one and one-to-many label assignments during training, violating DETR's one-to-one matching principle. To address these challenges, we propose DS-Det, a more efficient detector capable of detecting a flexible number of objects in images. Specifically, we reformulate and introduce a new unified Single-Query paradigm for decoder modeling, transforming the fixed-query into flexible. Furthermore, we propose a simplified decoder framework through attention disentangled learning: locating boxes with Cross-Attention (one-to-many process), deduplicating predictions with Self-Attention (one-to-one process), addressing "query ambiguity" and "ROT" issues directly, and enhancing decoder efficiency. We further introduce a unified PoCoo loss that leverages box size priors to prioritize query learning on hard samples such as small objects. Extensive experiments across five different backbone models on COCO2017 and WiderPerson datasets demonstrate the general effectiveness and superiority of DS-Det. The source codes are available at https://github.com/Med-Process/DS-Det/.

Guiping Cao, Wenjian Huang, Xiangyuan Lan et al.

Small Object Detection (SOD) poses significant challenges due to limited information and the model's low class prediction score. While Transformer-based detectors have shown promising performance, their potential for SOD remains largely unexplored. In typical DETR-like frameworks, the CNN backbone network, specialized in aggregating local information, struggles to capture the necessary contextual information for SOD. The multiple attention layers in the Transformer Encoder face difficulties in effectively attending to small objects and can also lead to blurring of features. Furthermore, the model's lower class prediction score of small objects compared to large objects further increases the difficulty of SOD. To address these challenges, we introduce a novel approach called Cross-DINO. This approach incorporates the deep MLP network to aggregate initial feature representations with both short and long range information for SOD. Then, a new Cross Coding Twice Module (CCTM) is applied to integrate these initial representations to the Transformer Encoder feature, enhancing the details of small objects. Additionally, we introduce a new kind of soft label named Category-Size (CS), integrating the Category and Size of objects. By treating CS as new ground truth, we propose a new loss function called Boost Loss to improve the class prediction score of the model. Extensive experimental results on COCO, WiderPerson, VisDrone, AI-TOD, and SODA-D datasets demonstrate that Cross-DINO efficiently improves the performance of DETR-like models on SOD. Specifically, our model achieves 36.4% APs on COCO for SOD with only 45M parameters, outperforming the DINO by +4.4% APS (36.4% vs. 32.0%) with fewer parameters and FLOPs, under 12 epochs training setting. The source codes will be available at https://github.com/Med-Process/Cross-DINO.

Yushuai Sun, Zikun Zhou, Dongmei Jiang et al.

Asymmetric retrieval is a typical scenario in real-world retrieval systems, where compatible models of varying capacities are deployed on platforms with different resource configurations. Existing methods generally train pre-defined networks or subnetworks with capacities specifically designed for pre-determined platforms, using compatible learning. Nevertheless, these methods suffer from limited flexibility for multi-platform deployment. For example, when introducing a new platform into the retrieval systems, developers have to train an additional model at an appropriate capacity that is compatible with existing models via backward-compatible learning. In this paper, we propose a Prunable Network with self-compatibility, which allows developers to generate compatible subnetworks at any desired capacity through post-training pruning. Thus it allows the creation of a sparse subnetwork matching the resources of the new platform without additional training. Specifically, we optimize both the architecture and weight of subnetworks at different capacities within a dense network in compatible learning. We also design a conflict-aware gradient integration scheme to handle the gradient conflicts between the dense network and subnetworks during compatible learning. Extensive experiments on diverse benchmarks and visual backbones demonstrate the effectiveness of our method. Our code and model are available at https://github.com/Bunny-Black/PrunNet.

Zaijing Li, Gongwei Chen, Rui Shao et al.

Large Language Models (LLMs) have shown remarkable performance in various emotion recognition tasks, thereby piquing the research community's curiosity for exploring their potential in emotional intelligence. However, several issues in the field of emotional generation tasks remain unresolved, including human preference alignment and emotional generation assessment. In this paper, we propose the Emotional Chain-of-Thought (ECoT), a plug-and-play prompting method that enhances the performance of LLMs on various emotional generation tasks by aligning with human emotional intelligence guidelines. To assess the reliability of ECoT, we propose an automated model-based evaluation method called Emotional Generation Score (EGS). EGS incorporates Goleman's Emotional Intelligence Theory as a consensus of human experts, providing a new perspective on the evaluation of emotional generation tasks. Extensive experimental results demonstrate the effectiveness of ECoT and EGS. Further, we discuss the promise of LLMs in the field of emotional intelligence and present key insights into the LLMs with the ECoT in emotional generation tasks.

Weikang Meng, Yadan Luo, Xin Li et al.

Linear attention has emerged as a promising alternative to softmax-based attention, leveraging kernelized feature maps to reduce complexity from quadratic to linear in sequence length. However, the non-negative constraint on feature maps and the relaxed exponential function used in approximation lead to significant information loss compared to the original query-key dot products, resulting in less discriminative attention maps with higher entropy. To address the missing interactions driven by negative values in query-key pairs, we propose a polarity-aware linear attention mechanism that explicitly models both same-signed and opposite-signed query-key interactions, ensuring comprehensive coverage of relational information. Furthermore, to restore the spiky properties of attention maps, we provide a theoretical analysis proving the existence of a class of element-wise functions (with positive first and second derivatives) that can reduce entropy in the attention distribution. For simplicity, and recognizing the distinct contributions of each dimension, we employ a learnable power function for rescaling, allowing strong and weak attention signals to be effectively separated. Extensive experiments demonstrate that the proposed PolaFormer improves performance on various vision tasks, enhancing both expressiveness and efficiency by up to 4.6%.

Zaijing Li, Yuquan Xie, Rui Shao et al.

Building an agent that can mimic human behavior patterns to accomplish various open-world tasks is a long-term goal. To enable agents to effectively learn behavioral patterns across diverse tasks, a key challenge lies in modeling the intricate relationships among observations, actions, and language. To this end, we propose Optimus-2, a novel Minecraft agent that incorporates a Multimodal Large Language Model (MLLM) for high-level planning, alongside a Goal-Observation-Action Conditioned Policy (GOAP) for low-level control. GOAP contains (1) an Action-guided Behavior Encoder that models causal relationships between observations and actions at each timestep, then dynamically interacts with the historical observation-action sequence, consolidating it into fixed-length behavior tokens, and (2) an MLLM that aligns behavior tokens with open-ended language instructions to predict actions auto-regressively. Moreover, we introduce a high-quality Minecraft Goal-Observation-Action (MGOA)} dataset, which contains 25,000 videos across 8 atomic tasks, providing about 30M goal-observation-action pairs. The automated construction method, along with the MGOA dataset, can contribute to the community's efforts to train Minecraft agents. Extensive experimental results demonstrate that Optimus-2 exhibits superior performance across atomic tasks, long-horizon tasks, and open-ended instruction tasks in Minecraft. Please see the project page at https://cybertronagent.github.io/Optimus-2.github.io/.

Yan Li, Yifei Xing, Xiangyuan Lan et al.

Cross-modal alignment is crucial for multimodal representation fusion due to the inherent heterogeneity between modalities. While Transformer-based methods have shown promising results in modeling inter-modal relationships, their quadratic computational complexity limits their applicability to long-sequence or large-scale data. Although recent Mamba-based approaches achieve linear complexity, their sequential scanning mechanism poses fundamental challenges in comprehensively modeling cross-modal relationships. To address this limitation, we propose AlignMamba, an efficient and effective method for multimodal fusion. Specifically, grounded in Optimal Transport, we introduce a local cross-modal alignment module that explicitly learns token-level correspondences between different modalities. Moreover, we propose a global cross-modal alignment loss based on Maximum Mean Discrepancy to implicitly enforce the consistency between different modal distributions. Finally, the unimodal representations after local and global alignment are passed to the Mamba backbone for further cross-modal interaction and multimodal fusion. Extensive experiments on complete and incomplete multimodal fusion tasks demonstrate the effectiveness and efficiency of the proposed method.

Songjun Tu, Jiahao Lin, Xiangyu Tian et al.

Recent advancements in post-training methodologies for large language models (LLMs) have highlighted reinforcement learning (RL) as a critical component for enhancing reasoning. However, the substantial computational costs associated with RL-based approaches have led to growing interest in alternative paradigms, such as Direct Preference Optimization (DPO). In this study, we investigate the effectiveness of DPO in facilitating self-improvement for LLMs through iterative preference-based learning. We demonstrate that a single round of DPO with coarse filtering significantly enhances mathematical reasoning performance, particularly for strong base model. Furthermore, we design an iterative enhancement framework for both the generator and the reward model (RM), enabling their mutual improvement through online interaction across multiple rounds of DPO. Finally, with simple verifiable rewards, our model DPO-VP achieves RL-level performance with significantly lower computational overhead. These findings highlight DPO as a scalable and cost-effective alternative to RL, offering a practical solution for enhancing LLM reasoning in resource-constrained situations.

Zheng Chong, Wenqing Zhang, Shiyue Zhang et al.

Virtual try-on (VTON) technology has gained attention due to its potential to transform online retail by enabling realistic clothing visualization of images and videos. However, most existing methods struggle to achieve high-quality results across image and video try-on tasks, especially in long video scenarios. In this work, we introduce CatV2TON, a simple and effective vision-based virtual try-on (V2TON) method that supports both image and video try-on tasks with a single diffusion transformer model. By temporally concatenating garment and person inputs and training on a mix of image and video datasets, CatV2TON achieves robust try-on performance across static and dynamic settings. For efficient long-video generation, we propose an overlapping clip-based inference strategy that uses sequential frame guidance and Adaptive Clip Normalization (AdaCN) to maintain temporal consistency with reduced resource demands. We also present ViViD-S, a refined video try-on dataset, achieved by filtering back-facing frames and applying 3D mask smoothing for enhanced temporal consistency. Comprehensive experiments demonstrate that CatV2TON outperforms existing methods in both image and video try-on tasks, offering a versatile and reliable solution for realistic virtual try-ons across diverse scenarios.

Zhehan Kan, Ce Zhang, Zihan Liao et al.

Large Vision-Language Model (LVLM) systems have demonstrated impressive vision-language reasoning capabilities but suffer from pervasive and severe hallucination issues, posing significant risks in critical domains such as healthcare and autonomous systems. Despite previous efforts to mitigate hallucinations, a persistent issue remains: visual defect from vision-language misalignment, creating a bottleneck in visual processing capacity. To address this challenge, we develop Complementary Adaptive Token-level Contrastive Decoding to Mitigate Hallucinations in LVLMs (CATCH), based on the Information Bottleneck theory. CATCH introduces Complementary Visual Decoupling (CVD) for visual information separation, Non-Visual Screening (NVS) for hallucination detection, and Adaptive Token-level Contrastive Decoding (ATCD) for hallucination mitigation. CATCH addresses issues related to visual defects that cause diminished fine-grained feature perception and cumulative hallucinations in open-ended scenarios. It is applicable to various visual question-answering tasks without requiring any specific data or prior knowledge, and generalizes robustly to new tasks without additional training, opening new possibilities for advancing LVLM in various challenging applications.

Yong Dai, Xiaopeng Hong, Yabin Wang et al.

Contemporary continual learning approaches typically select prompts from a pool, which function as supplementary inputs to a pre-trained model. However, this strategy is hindered by the inherent noise of its selection approach when handling increasing tasks. In response to these challenges, we reformulate the prompting approach for continual learning and propose the prompt customization (PC) method. PC mainly comprises a prompt generation module (PGM) and a prompt modulation module (PMM). In contrast to conventional methods that employ hard prompt selection, PGM assigns different coefficients to prompts from a fixed-sized pool of prompts and generates tailored prompts. Moreover, PMM further modulates the prompts by adaptively assigning weights according to the correlations between input data and corresponding prompts. We evaluate our method on four benchmark datasets for three diverse settings, including the class, domain, and task-agnostic incremental learning tasks. Experimental results demonstrate consistent improvement (by up to 16.2\%), yielded by the proposed method, over the state-of-the-art (SOTA) techniques.

Wenyun Li, Zheng Zhang, Xiangyuan Lan et al.

Traditional adversarial attacks typically produce adversarial examples under norm-constrained conditions, whereas unrestricted adversarial examples are free-form with semantically meaningful perturbations. Current unrestricted adversarial impersonation attacks exhibit limited control over adversarial face attributes and often suffer from low transferability. In this paper, we propose a novel Text Controlled Attribute Attack (TCA$^2$) to generate photorealistic adversarial impersonation faces guided by natural language. Specifically, the category-level personal softmax vector is employed to precisely guide the impersonation attacks. Additionally, we propose both data and model augmentation strategies to achieve transferable attacks on unknown target models. Finally, a generative model, \textit{i.e}, Style-GAN, is utilized to synthesize impersonated faces with desired attributes. Extensive experiments on two high-resolution face recognition datasets validate that our TCA$^2$ method can generate natural text-guided adversarial impersonation faces with high transferability. We also evaluate our method on real-world face recognition systems, \textit{i.e}, Face++ and Aliyun, further demonstrating the practical potential of our approach.

Yuquan Xie, Zaijing Li, Rui Shao et al.

Recent efforts to leverage the Multi-modal Large Language Model (MLLM) as GUI agents have yielded promising outcomes. However, these agents still struggle with long-horizon tasks in online environments, primarily due to insufficient knowledge and the inherent gap between offline and online domains. In this paper, inspired by how humans generalize knowledge in open-ended environments, we propose a Hierarchical Multimodal Skills (HMS) module to tackle the issue of insufficient knowledge. It progressively abstracts trajectories into execution skills, core skills, and ultimately meta-skills, providing a hierarchical knowledge structure for long-horizon task planning. To bridge the domain gap, we propose the Skill-Augmented Monte Carlo Tree Search (SA-MCTS) algorithm, which efficiently leverages skills acquired in offline environments to reduce the action search space during online tree exploration. Building on HMS, we propose Mirage-1, a multimodal, cross-platform, plug-and-play GUI agent. To validate the performance of Mirage-1 in real-world long-horizon scenarios, we constructed a new benchmark, AndroidLH. Experimental results show that Mirage-1 outperforms previous agents by 32\%, 19\%, 15\%, and 79\% on AndroidWorld, MobileMiniWob++, Mind2Web-Live, and AndroidLH, respectively. Project page: https://cybertronagent.github.io/Mirage-1.github.io/

Zaijing Li, Bing Hu, Rui Shao et al.

Hierarchical Vision-Language-Action (VLA) models have rapidly become a dominant paradigm for robotic manipulation. It typically comprising a Vision-Language backbone for perception and understanding, together with a generative policy for action generation. However, its performance is increasingly bottlenecked by the action generation proceess. (i) Low inference efficiency. A pronounced distributional gap between isotropic noise priors and target action distributions, which increases denoising steps and the incidence of infeasible samples. (ii) Poor robustness. Existing policies condition solely on the current observation, neglecting the constraint of history sequence and thus lacking awareness of task progress and temporal consistency. To address these issues, we introduce OptimusVLA, a dual-memory VLA framework with Global Prior Memory (GPM) and Local Consistency Memory (LCM). GPM replaces Gaussian noise with task-level priors retrieved from semantically similar trajectories, thereby shortening the generative path and reducing the umber of function evaluations (NFE). LCM dynamically models executed action sequence to infer task progress and injects a learned consistency constraint that enforces temporal coherence and smoothness of trajectory. Across three simulation benchmarks, OptimusVLA consistently outperforms strong baselines: it achieves 98.6% average success rate on LIBERO, improves over pi_0 by 13.5% on CALVIN, and attains 38% average success rate on RoboTwin 2.0 Hard. In Real-World evaluation, OptimusVLA ranks best on Generalization and Long-horizon suites, surpassing pi_0 by 42.9% and 52.4%, respectively, while delivering 2.9x inference speedup.

Keliang Li, Hongze Shen, Hao Shi et al.

The aspiration for artificial general intelligence, fueled by the rapid progress of multimodal models, demands human-comparable performance across diverse environments. We propose HumanPCR, an evaluation suite for probing MLLMs' capacity about human-related visual contexts across three hierarchical levels: Perception, Comprehension, and Reasoning (denoted by Human-P, Human-C, and Human-R, respectively). Human-P and Human-C feature over 6,000 human-verified multiple choice questions, assessing massive tasks of 9 dimensions, including but not limited to essential skills frequently overlooked by existing benchmarks. Human-R offers a challenging manually curated video reasoning test that requires integrating multiple visual evidences, proactively extracting context beyond question cues, and applying human-like expertise. Each question includes human-annotated Chain-of-Thought (CoT) rationales with key visual evidence to support further research. Extensive evaluations on over 30 state-of-the-art models exhibit significant challenges in human-centric visual understanding, particularly in tasks involving detailed space perception, temporal understanding, and mind modeling. Moreover, analysis of Human-R reveals the struggle of models in extracting essential proactive visual evidence from diverse human scenes and their faulty reliance on query-guided retrieval. Even with advanced techniques like scaling visual contexts and test-time thinking yield only limited benefits. We hope HumanPCR and our findings will advance the development, evaluation, and human-centric application of multimodal models.

Zaijing Li, Yuquan Xie, Rui Shao et al.

Recently, agents based on multimodal large language models (MLLMs) have achieved remarkable progress across various domains. However, building a generalist agent with capabilities such as perception, planning, action, grounding, and reflection in open-world environments like Minecraft remains challenges: insufficient domain-specific data, interference among heterogeneous tasks, and visual diversity in open-world settings. In this paper, we address these challenges through three key contributions. 1) We propose a knowledge-enhanced data generation pipeline to provide scalable and high-quality training data for agent development. 2) To mitigate interference among heterogeneous tasks, we introduce a Mixture-of-Experts (MoE) architecture with task-level routing. 3) We develop a Multimodal Reasoning-Augmented Reinforcement Learning approach to enhance the agent's reasoning ability for visual diversity in Minecraft. Built upon these innovations, we present Optimus-3, a general-purpose agent for Minecraft. Extensive experimental results demonstrate that Optimus-3 surpasses both generalist multimodal large language models and existing state-of-the-art agents across a wide range of tasks in the Minecraft environment. Project page: https://cybertronagent.github.io/Optimus-3.github.io/

Yan Li, Yong Zhao, Xiaohan Xia et al.

The key to facial expression recognition is to learn discriminative spatial-temporal representations that embed facial expression dynamics. Previous studies predominantly rely on pre-trained Convolutional Neural Networks (CNNs) to learn facial appearance representations, overlooking the relationships between facial regions. To address this issue, this paper presents an Appearance- and Relation-aware Parallel Graph attention fusion Network (ARPGNet) to learn mutually enhanced spatial-temporal representations of appearance and relation information. Specifically, we construct a facial region relation graph and leverage the graph attention mechanism to model the relationships between facial regions. The resulting relational representation sequences, along with CNN-based appearance representation sequences, are then fed into a parallel graph attention fusion module for mutual interaction and enhancement. This module simultaneously explores the complementarity between different representation sequences and the temporal dynamics within each sequence. Experimental results on three facial expression recognition datasets demonstrate that the proposed ARPGNet outperforms or is comparable to state-of-the-art methods.

Wenyun Li, Zheng Zhang, Dongmei Jiang et al.

Large language models (LLMs) have garnered significant interest in AI community. Despite their impressive generation capabilities, they have been found to produce misleading or fabricated information, a phenomenon known as hallucinations. Consequently, hallucination detection has become critical to ensure the reliability of LLM-generated content. One primary challenge in hallucination detection is the scarcity of well-labeled datasets containing both truthful and hallucinated outputs. To address this issue, we introduce Prompt-guided data Augmented haLlucination dEtection (PALE), a novel framework that leverages prompt-guided responses from LLMs as data augmentation for hallucination detection. This strategy can generate both truthful and hallucinated data under prompt guidance at a relatively low cost. To more effectively evaluate the truthfulness of the sparse intermediate embeddings produced by LLMs, we introduce an estimation metric called the Contrastive Mahalanobis Score (CM Score). This score is based on modeling the distributions of truthful and hallucinated data in the activation space. CM Score employs a matrix decomposition approach to more accurately capture the underlying structure of these distributions. Importantly, our framework does not require additional human annotations, offering strong generalizability and practicality for real-world applications. Extensive experiments demonstrate that PALE achieves superior hallucination detection performance, outperforming the competitive baseline by a significant margin of 6.55%.

Songjun Tu, Qichao Zhang, Jingbo Sun et al.

While multimodal large language models excel at tasks that integrate visual perception with symbolic reasoning, their performance is often undermined by a critical vulnerability: perception-induced errors that propagate through the reasoning chain. Current reinforcement learning (RL) fine-tuning methods, while enhancing reasoning abilities, largely fail to address the underlying misalignment between visual grounding and the subsequent reasoning process. To address this challenge, we propose \textbf{Caption-Regularized Policy Optimization (CapPO)}, a novel RL framework that explicitly enforces perceptual consistency during policy optimization. CapPO integrates two key mechanisms: (1) a caption-based consistency regularization, which minimizes the divergence between responses conditioned on raw images and those conditioned on captions, thereby anchoring reasoning to semantically faithful visual content; and (2) a KL-weighted advantage estimation scheme, which adaptively scales reinforcement signals to strengthen perceptually consistent trajectories while suppressing spurious correlations. Extensive experiments on five math-focused and five general reasoning benchmarks demonstrate that CapPO achieves competitive performance, yielding gains of +6.0% accuracy on math-related tasks and +2.4% on general reasoning tasks over the base Qwen2.5-VL-7B model. Moreover, ablation studies further confirm the effectiveness of each component, while error analysis reveals that CapPO significantly reduces perception-related mistakes compared with baselines. Overall, CapPO provides a simple yet effective framework for improving multimodal reasoning.

Zheng Chong, Yanwei Lei, Shiyue Zhang et al.

Despite its great potential, virtual try-on technology is hindered from real-world application by two major challenges: the inability of current methods to support multi-reference outfit compositions (including garments and accessories), and their significant inefficiency caused by the redundant re-computation of reference features in each denoising step. To address these challenges, we propose FastFit, a high-speed multi-reference virtual try-on framework based on a novel cacheable diffusion architecture. By employing a Semi-Attention mechanism and substituting traditional timestep embeddings with class embeddings for reference items, our model fully decouples reference feature encoding from the denoising process with negligible parameter overhead. This allows reference features to be computed only once and losslessly reused across all steps, fundamentally breaking the efficiency bottleneck and achieving an average 3.5x speedup over comparable methods. Furthermore, to facilitate research on complex, multi-reference virtual try-on, we introduce DressCode-MR, a new large-scale dataset. It comprises 28,179 sets of high-quality, paired images covering five key categories (tops, bottoms, dresses, shoes, and bags), constructed through a pipeline of expert models and human feedback refinement. Extensive experiments on the VITON-HD, DressCode, and our DressCode-MR datasets show that FastFit surpasses state-of-the-art methods on key fidelity metrics while offering its significant advantage in inference efficiency.

Yaguang Song, Xiaoshan Yang, Dongmei Jiang et al.

Incremental learning aims to enable models to continuously acquire knowledge from evolving data streams while preserving previously learned capabilities. While current research predominantly focuses on unimodal incremental learning and multimodal incremental learning where the modalities are consistent, real-world scenarios often present data from entirely new modalities, posing additional challenges. This paper investigates the feasibility of developing a unified model capable of incremental learning across continuously evolving modal sequences. To this end, we introduce a novel paradigm called Modality Incremental Learning (MIL), where each learning stage involves data from distinct modalities. To address this task, we propose a novel framework named Harmony, designed to achieve modal alignment and knowledge retention, enabling the model to reduce the modal discrepancy and learn from a sequence of distinct modalities, ultimately completing tasks across multiple modalities within a unified framework. Our approach introduces the adaptive compatible feature modulation and cumulative modal bridging. Through constructing historical modal features and performing modal knowledge accumulation and alignment, the proposed components collaboratively bridge modal differences and maintain knowledge retention, even with solely unimodal data available at each learning phase.These components work in concert to establish effective modality connections and maintain knowledge retention, even when only unimodal data is available at each learning stage. Extensive experiments on the MIL task demonstrate that our proposed method significantly outperforms existing incremental learning methods, validating its effectiveness in MIL scenarios.

Jiyao Liu, Yanxi Zhao, Hao Wu et al.

Recognizing the feelings of human beings plays a critical role in our daily communication. Neuroscience has demonstrated that different emotion states present different degrees of activation in different brain regions, EEG frequency bands and temporal stamps. In this paper, we propose a novel structure to explore the informative EEG features for emotion recognition. The proposed module, denoted by PST-Attention, consists of Positional, Spectral and Temporal Attention modules to explore more discriminative EEG features. Specifically, the Positional Attention module is to capture the activate regions stimulated by different emotions in the spatial dimension. The Spectral and Temporal Attention modules assign the weights of different frequency bands and temporal slices respectively. Our method is adaptive as well as efficient which can be fit into 3D Convolutional Neural Networks (3D-CNN) as a plug-in module. We conduct experiments on two real-world datasets. 3D-CNN combined with our module achieves promising results and demonstrate that the PST-Attention is able to capture stable patterns for emotion recognition from EEG.

Congqi Cao, Yue Lu, Yifan Zhang et al.

Contextual information plays an important role in action recognition. Local operations have difficulty to model the relation between two elements with a long-distance interval. However, directly modeling the contextual information between any two points brings huge cost in computation and memory, especially for action recognition, where there is an additional temporal dimension. Inspired from 2D criss-cross attention used in segmentation task, we propose a recurrent 3D criss-cross attention (RCCA-3D) module to model the dense long-range spatiotemporal contextual information in video for action recognition. The global context is factorized into sparse relation maps. We model the relationship between points in the same line along the direction of horizon, vertical and depth at each time, which forms a 3D criss-cross structure, and duplicate the same operation with recurrent mechanism to transmit the relation between points in a line to a plane finally to the whole spatiotemporal space. Compared with the non-local method, the proposed RCCA-3D module reduces the number of parameters and FLOPs by 25% and 30% for video context modeling. We evaluate the performance of RCCA-3D with two latest action recognition networks on three datasets and make a thorough analysis of the architecture, obtaining the optimal way to factorize and fuse the relation maps. Comparisons with other state-of-the-art methods demonstrate the effectiveness and efficiency of our model.

Meshia Cédric Oveneke, Mitchel Aliosha-Perez, Yong Zhao et al.

The omnipresence of deep learning architectures such as deep convolutional neural networks (CNN)s is fueled by the synergistic combination of ever-increasing labeled datasets and specialized hardware. Despite the indisputable success, the reliance on huge amounts of labeled data and specialized hardware can be a limiting factor when approaching new applications. To help alleviating these limitations, we propose an efficient learning strategy for layer-wise unsupervised training of deep CNNs on conventional hardware in acceptable time. Our proposed strategy consists of randomly convexifying the reconstruction contractive auto-encoding (RCAE) learning objective and solving the resulting large-scale convex minimization problem in the frequency domain via coordinate descent (CD). The main advantages of our proposed learning strategy are: (1) single tunable optimization parameter; (2) fast and guaranteed convergence; (3) possibilities for full parallelization. Numerical experiments show that our proposed learning strategy scales (in the worst case) linearly with image size, number of filters and filter size.