Recently, diffusion models have shown strong potential for 3D point cloud generation, but capturing point distributions remains difficult due to data sparsity and irregularity. We propose a method that integrates Optimal Transport with diffusion models. Specifically, point clouds are first mapped into a voxel grid via Optimal Transport, then processed with diffusion’s forward and reverse stages. To improve the quality of generated point clouds, we design a point cloud–specific image-level loss and adopt a weighted objective with an attention denoising network. Experiments on ShapeNet demonstrate that our approach generates high-quality, diverse point clouds with improved efficiency compared to existing diffusion-based methods.

This paper investigates the strong correlation between interaction methods and scenario requirements within the specialized context of children’s education. It proposes an innovative integration of multimodal interaction techniques to bridge augmented reality (AR), garment design, and educational content. A comprehensive solution based on marker-based visual tracking is developed to support this integration. Key technical challenges such as real-time marker extraction, frame sequence jitter, logo pattern recognition, and texture mapping are addressed. The system allows dynamic combination and presentation of garment card components, enabling interactive learning scenarios. Features such as 360-degree preview and a rotatable mannequin mode are implemented to enhance natural interaction, reduce cost, and improve versatility. Experimental evaluations demonstrate an average processing time of 37 milliseconds per frame and a frame rate of 27 frames per second on mainstream smartphones, meeting real-time application demands. Furthermore, learning effectiveness experiments show that the proposed system significantly enhances children’s learning outcomes, engagement, and overall performance. This work highlights the potential of multimodal AR interactions to transform traditional educational experiences through culturally enriched, visually engaging, and user-centered design.

With the advancement of 3D Gaussian Splatting, dynamic scene reconstruction has achieved significant progress. However, previous approaches typically model deformations independently for each frame, lacking spatiotemporal consistency and limiting reconstruction quality. In addition, existing methods still face challenges in monocular settings, particularly when handling complex motions. To address these issues, we propose a novel framework for monocular dynamic reconstruction. Our framework captures motion trends of Gaussian primitives from adjacent frames and leverages them to enhance both Gaussian and temporal feature representations via an attention mechanism. Moreover, we further decompose the overall deformations into coarse and fine components to better model complex motions. Finally, we introduce geometric priors from depth and normal maps to regularize the optimization process. Extensive experiments on various real-world datasets demonstrate that our method achieves state-of-the-art reconstruction quality. Project page: https://sffxh.github.io/MT-GS.

In many real‑world intelligent decision‑support systems, the ability to obtain compact yet discriminative multi-scale representations is crucial for ensuring reliable and efficient downstream performance. Motivated by these practical demands, multi-scale representation learning has gained increasing attention in recent years. Multi-scale representation learning plays a pivotal role in downstream tasks, yet it often suffers from semantic redundancy and noise interference, especially in cluttered or complex scenes. In this paper, we rethink the conventional paradigm from an information-theoretic perspective and propose a novel Hierarchical Information Bottleneck (HIB) formulation tailored for multi-scale architectures. Building upon this theory, we design HIB-MSRL, a principled representation learning framework that shifts the focus from conventional feature fusion to optimization-driven information compression and selective semantic retention. HIB-MSRL enables layer-wise redundancy suppression while preserving label-related information, thereby enhancing robustness and generalization. We further implement the HIB objective via variational inference, providing a principled and tractable optimization target for fine-grained information control. Extensive experiments across multiple benchmarks demonstrate that HIB-MSRL consistently outperforms state-of-the-art methods in both accuracy and robustness, particularly in cluttered or noisy environments. Our work offers new insights into multi-scale representation learning from an information-theoretic perspective.

In the evolving nexus of neuroscience and robotics, the symbiotic fusion of electroencephalography (EEG) and electromyography (EMG) is emerging as a paradigm-shifting avenue for enhancing human-machine interfaces. While EEG, which captures the subtle electrical nuances of the brain, offers a potent channel for nuanced brain-machine communication, EMG serves as a bridge, converting neuromuscular intentions into actionable directives for robotic apparatuses. This review highlights the current methodologies in which EEG and EMG not only function in silos but also converge harmoniously to dictate robotic control. By delving deeper into this, the intricate synergy between cognitive processes, muscular responses, and machine actions can be unraveled. Subsequently, the discourse also navigates through the myriad challenges encountered in realizing real-time, seamless integration of these bio-signals with robotics and the innovative solutions poised to address them. The aim is to provide a comprehensive understanding of the interplay between neuroscience and robotics. This insight will help drive breakthroughs in adaptive human-machine collaboration.

Transformer-based architectures have demonstrated strong performance in 3D point cloud understanding, yet many existing methods generate attention weights mainly from semantic feature similarity. In deep networks, feature-centric attention may become less selective as point features are progressively smoothed, a behavior associated with feature homogenization and rank collapse, which can weaken the structural discrimination of local aggregation.We propose PointGP, a geometry-primed framework that uses rectified local geometric topology as the primary cue for attention generation. PointGP introduces a Semantic-Guided Manifold Rectifier to predict feature-conditioned local coordinate offsets, and a Dual-Stream Geometric Kernel to compute attention logits from both raw and rectified geometric cues. By reducing reliance on explicit query-key feature matching while implicitly incorporating semantic guidance through geometric rectification, PointGP provides an effective and efficient mechanism for local point aggregation.Experiments across five benchmarks covering classification, part segmentation, and indoor scene segmentation show that PointGP achieves competitive accuracy with strong parameter and computational efficiency compared with representative strong baselines.

Reconstructing dynamic objects from monocular RGB-D video is critical for advancing 3D vision applications and enhancing user experience. However, monocular RGB-D video provides limited 3D observations, making the reconstruction of unobserved regions highly under-constrained. Despite recent advances that combine neural implicit surfaces with diffusion models, the inherent limitations of implicit representations and the lack of effective guidance in diffusion priors lead to blurry appearance and inaccurate geometry in dynamic object reconstruction. To address the issue, we present MGD, which leverages scene-adaptive diffusion priors and Mesh-guided Gaussians for realistic rendering and geometrically accurate reconstruction of dynamic objects, including unobserved regions. The dynamic 3D objects reconstructed by MGD are represented using our proposed Mesh-guided Gaussians, which leverage global and local Gaussians to capture large-scale deformations and fine-grained appearance details, respectively. Additionally, in order to utilize depth information, we integrate a depth ControlNet into the diffusion model and conduct scene-adaptive fine-tuning. We design a self-generated image-pair strategy to produce image pairs used for fine-tuning. Extensive experiments demonstrate that MGD achieves state-of-the-art performance in both high-fidelity reconstruction and structural completeness, while maintaining real-time efficiency during training and rendering.

EEG foundation models (EFMs) achieve strong cross-subject and cross-task generalization through large-scale pretraining and downstream fine-tuning. Through empirical analysis, we observe that (i) task-adapted EFMs provide strong decoding performance but incur substantial overhead when retained as inference backbones, making knowledge distillation a natural route for optimizing compact students; and (ii) direct distillation from a fixed teacher representation underutilizes EFM knowledge, as task-discriminative information is distributed across intermediate layers rather than concentrated in the final layer. These observations motivate DLink (Distilling Layer-wise and Dominant Knowledge), a spectrally guided distillation framework with input-conditioned layer routing for transferring EFM knowledge into compact students. DLink uses a lightweight router to aggregate teacher layers for each input, and aligns magnitude and phase spectra to mitigate compression-induced spectral distortion in learned representations. The routed teacher knowledge is internalized by a project-then-compress student; the teacher and router are used only during training. Experiments on four EEG benchmarks show that DLink improves matched compact students and remains competitive with lightweight baselines, narrowing the gap to fine-tuned EFMs while substantially reducing parameters, FLOPs, and CPU-only inference latency.

Multimodal fake news video detection is a crucial research direction for maintaining the credibility of online information. Existing studies primarily verify content authenticity by constructing multimodal feature fusion representations or utilizing pre-trained language models to analyze video-text consistency. However, these methods still face the following limitations: (1) lacking cross-instance global semantic correlations, making it difficult to effectively utilize historical associative evidence to verify the current video; (2) semantic discrepancies across domains hinder the transfer of general knowledge, lacking the guidance of domain-specific expert knowledge. To this end, we propose a novel Retrieval-Augmented Semantic Reasoning (RASR) framework. First, a Cross-instance Semantic Parser and Retriever (CSPR) deconstructs the video into high-level semantic primitives and retrieves relevant associative evidence from a dynamic memory bank. Subsequently, a Domain-Guided Multimodal Reasoning (DGMP) module incorporates domain priors to drive an expert multimodal large language model in generating domain-aware, in-depth analysis reports. Finally, a Multi-View Feature Decoupling and Fusion (MVDFF) module integrates multi-dimensional features through an adaptive gating mechanism to achieve robust authenticity determination. Extensive experiments on the FakeSV and FakeTT datasets demonstrate that RASR significantly outperforms state-of-the-art baselines, achieves superior cross-domain generalization, and improves the overall detection accuracy by up to 0.93

Automatic misinformation detection performs well when deception is visible in what an article explicitly states. However, some misinformation articles remain locally coherent and only become misleading once compared with contemporaneous reports that supply background facts the article omits. We study this omission-relevant setting and observe that current omission-aware approaches typically either attach retrieved context as auxiliary evidence or infer a categorical omission signal, leaving the specific missing fact implicit. We propose Latent Causal Void (LCV), a retrieval-guided detector that explicitly reconstructs the missing fact for each target sentence and uses it as a textual cross-source relation in graph reasoning. Concretely, LCV retrieves temporally aligned context articles, asks a frozen instruction-tuned large language model to generate a short missing-context description for each sentence–article pair, and feeds the resulting relation text into a heterograph over target sentences and context articles. On the bilingual benchmark of Sheng et al., LCV improves over the strongest omission-aware baseline by 2.56 and 2.84 macro-F1 points on the English and Chinese splits, respectively. The results indicate that modeling the missing cross-source fact itself, rather than only attaching retrieved evidence or predicting an omission signal, is a useful representation for omission-aware misinformation detection.

In this paper, we present a simple yet effective method RBF-MAT, for computing medial axis transform (MAT) from point cloud using radial basis functions (RBFs), where the surface is represented as the zero-level set of an interpolating function composed of a linear combination of RBFs. Firstly, we propose a new strategy for selecting the initial medial spheres based on the Voronoi vertices computed from the input points while preserving necessary geometric characteristics. Then the centers and radii of the generated medial spheres are iteratively optimized by minimizing the RBF-based surface reconstruction error. Additionally, the connectivity of the refined medial spheres is constructed as the dual of the restricted power diagram for these spheres. Experimental results across diverse 3D shapes demonstrate our method's efficacy in capturing global structural attributes and local geometric intricacies, with our connectivity approach surpassing existing methods. Besides, the experimental results show that the MATs computed with our method better approximate the point cloud surface than state-of-the-art methods.

Graph Neural Networks (GNNs) have recently seen extensive Collaborative Filtering (CF) applications. However, noisy interactions are usually contained in the original user-item interactions. In existing research, the influence of noisy interactions has not been simultaneously considered to be eliminated from both the embedding and sample spaces. To address this limitation, an innovative multi-view contrastive learning framework called Denoising Multi-View Graph Contrastive Learning (DMGCL) is proposed. In the sample space, denoising and augmented views are constructed based on structural and embedding similarity. In the embedding space, a complementary view is created to assist in correcting user interest modeling bias. Subsequently, contrastive learning is performed on these three views by DMGCL, denoising from both the sample space and the embedding space in a fine-grained manner. Additionally, random perturbations are introduced into the embeddings, and inter-layer contrastive learning is performed to achieve a more uniform embedding distribution. To demonstrate the performance of our proposed DMGCL, comprehensive experiments are conducted on four datasets from various domains. Evaluated on Tmall, Yelp, Gowalla, and Amazon-Book demonstrates that DMGCL outperforms the current state-of-the-art contrastive learning method with improvements of 1.58

Video stabilization aims to mitigate or eliminate the shake presented within video frames. Existing online video stabilization technologies rely on information from future frames, which may introduce a lag during real-time video stabilization. To surmount this hurdle, an online video stabilization model called OVST is proposed, which leverages solely historical video frames and enhanced by an attention mechanism. To simplify the complexity of model training and enhance robustness, a two-stage training strategy is proposed to decouple the fitting of real poses and the stabilization of virtual poses, and a hybrid stabilization loss with interframe soft constraints is designed, which effectively regulates the changes in camera poses between adjacent frames through interframe displacement, angular distortion, and cropping rate, thereby suppressing the distortion effects caused by excessive pose smoothing while balancing stability and cropping rate. Experiments demonstrate the superiority of the proposed OVST method over existing state-of-the-art techniques, achieving a stability metric of 0.8878 and a distortion metric of 0.9870.

Emotion Recognition in Conversation (ERC) is a challenging task due to the scarcity and dispersion of contextual information across utterances. Most existing methods attempt to integrate comprehensive information to enhance utterance semantics, which, however, also introduces noise and irrelevant content, misleading the model and limiting its potential in emotion recognition. To this end, we introduce the concept of Conversational Clique (ConvClique) and propose CC-ERC, a multi-view emotional knowledge extraction method designed to capture the most relevant emotional cues within the ConvClique from complementary perspectives and collaboratively predict utterance emotions. Specifically, CC-ERC comprises two modules: 1) the Utterance Spatial Relationship (USR) module, which predicts emotions by modeling structural correlations among utterances, and 2) the Emotion Temporal Relationship (ETR) module, which captures emotion sequence patterns to determine utterance emotions. These modules are integrated to obtain the final prediction, contributing to the robustness and accuracy of emotion recognition. The effectiveness of CC-ERC is validated on three widely used ERC datasets, evaluated in both online and offline settings. Compared to the state-of-the-art methods, CC-ERC achieves average improvements of 0.63% in accuracy and 0.94% in weighted F1 scores. Ablation studies further validate the significance of ConvClique-based knowledge extraction and demonstrate the effectiveness of the USR and ETR modules in modeling utterance structural correlations and emotion sequence patterns.

Graph Neural Networks (GNNs) have significantly advanced recommendation systems by modeling user-item interactions through bipartite graphs. However, real-world user-item interaction data are often sparse and noisy. Traditional bipartite graph modeling fails to capture higher-order relationships between users and items, limiting the ability of GNNs to learn high-quality node embeddings. While existing graph contrastive learning methods address data sparsity by partitioning nodes into positive and negative pairs, they also neglect these higher-order relationships, thus limiting the effectiveness of contrastive learning in recommendation systems. Furthermore, due to the inherent limitations of graph convolution, noise can propagate and amplify with increasing layers in deep graph convolutional networks. To address these challenges, Neighbor Enhancement and Embedding Perturbation for Graph Contrastive Learning (NPGCL) is proposed, which introduces two key modules - Relational Neighbor Enhancement Module and Collaborative Neighbor Enhancement Module - to capture higher-order relationships between homogeneous nodes and calculate interaction importance for noise suppression. Moreover, NPGCL employs an Embedding Perturbation Strategy and applies inter-layer contrastive learning to mitigate the noise impact caused by multi-layer graph convolutions. Experimental results demonstrate that NPGCL significantly improves performance across four publicly available datasets, with a notable enhancement in robustness, especially in noisy environments. Specifically, NPGCL achieves performance improvements of 1.77

Conversational query generation aims at producing search queries from dialogue histories, which are then used to retrieve relevant knowledge from a search engine to help knowledge- based dialogue systems. Trained to maximize the likelihood of gold queries, previous models suffer from the data hunger issue, and they tend to both drop important concepts from dialogue histories and generate irrelevant concepts at inference time. We attribute these issues to the over-association phenomenon where a large number of gold queries are indirectly related to the dialogue topics, because annotators may unconsciously perform reasoning with their background knowledge when generating these gold queries. We carefully analyze the negative effects of this phenomenon on pretrained Seq2seq query producers and then propose effective instance-level weighting strategies for training to mitigate these issues from multiple perspectives. Experiments on two benchmarks, Wizard-of-Internet and DuSinc, show that our strategies effectively alleviate the negative effects and lead to significant performance gains (2%similar to similar to 5% across automatic metrics and human evaluation). Further analysis shows that our model selects better concepts from dialogue histories and is 10 times more data efficient than the baseline.