Generative models have recently achieved excellent performance in the field of speech enhancement (SE), but their practical deployment is still hindered by heavy inference costs, including repeated function evaluations, expensive backbones, and inefficient sampling from generic priors. In this paper, we propose SCOPE (Sparsity-Constrained priOr for Progressive speech Enhancement), an efficient flow-matching framework designed to improve the quality–efficiency frontier of generative SE. SCOPE is motivated by two overlooked properties of SE: the stage-dependent nature of the denoising process and the inherent spectral sparsity of speech signals. To exploit these properties, we introduce Progressive Denoising Flow, which partitions the trajectory into specialized segments to decouple coarse structure recovery from fine harmonic refinement, and a Sparsity-Constrained Prior, which restricts the initial noise to spectral dominance regions to avoid unnecessary generation in silent areas. We further develop a Shared-Backbone Diffusion Transformer (DiT) that combines universal feature extraction with stage-specific specialization while reducing active inference costs. Extensive experiments on VoiceBank-DEMAND and DNS Challenge 2020 demonstrate that SCOPE achieves state-of-the-art-level enhancement quality using only four inference steps. Compared with VPIDM, SCOPE provides a 16.6× RTF speedup and reduces GFLOPs by over 250×. It also provides RTF speedups of 2.4× over UNIVERSE++.

Multi-modal knowledge graph completion (MMKGC) aims to infer missing factual triples by integrating structural information with rich multi-modal data such as text and images. While existing methods design various multi-modal fusion mechanisms, they largely overlook the inherent disparities in modality quality, as entity modalities across triples differ in informativeness, uncertainty, and noise, and they fail to address sample-specific modality uncertainty, ultimately resulting in suboptimal performance. To address this limitation, we propose MIMoE, a Modality-Interaction Enhanced Mixture-of-Experts framework with an uncertainty-aware router that adaptively integrates heterogeneous modalities. Specifically, MIMoE adopts a two-stage training paradigm: intra-modal pre-training to capture modality-specific semantics, followed by inter-modal post-training to model fine-grained cross-modal interactions with uncertainty-aware adaptive routing. Extensive experiments on three benchmarks demonstrate the effectiveness of MIMoE.

As the core tasks of spoken language understanding (SLU), the performance of intent classification and slot filling (ICSF) is often degraded by automatic speech recognition (ASR) errors. Existing methods typically depend on ASR internals or system-specific adaptation, limiting scalability. We propose DOMA, a model-agnostic framework that refines ASR transcripts for ICSF with diffusion language models (DLMs) and adaptive prior. DOMA generates multiple tokens in parallel at each decoding step and leverages the selectively preserved token sequence as a prior for the reverse process, enabling faster and more controllable refinement compared to autoregressive models. On benchmark datasets, DOMA improves ICSF performance by up to 3.2% relative to strong baselines, while reducing inference latency by 34.8% compared with standard DLM. Code is available at https://github.com/ICDM-UESTC/DOMA.

Memes are ubiquitous on social networks, but they often hide hateful content, as benign images or texts can become offensive when combined. Existing detection methods struggle with complex multimodal causal relationships, making it difficult to isolate the features that drive hateful intent. To address this challenge, we propose CAFT (Counter-factual Attribution & Feature Transformer), which constructs counterfactual sample spaces through controlled text interventions, leverages a pretrained multimodal encoder to capture feature discrepancies between anchor and counterfactual samples, and disentangles causal signals. These representations are further refined by consistency and contrastive losses, which suppress spurious correlations. Extensive experiments on standard meme benchmarks demonstrate that CAFT surpasses state-of-the-art methods, enhances robustness against adversarial edits, and yields interpretable token-level evidence for identifying causal factors underlying hateful memes.

Weight Averaging (WA) has emerged as a powerful technique for enhancing generalization by promoting convergence to a flat loss landscape, which correlates with stronger out-of-distribution performance. However, applying WA directly to multi-modal domain generalization (MMDG) is challenging: differences in optimization speed across modalities lead WA to overfit to faster-converging ones in early stages, suppressing the contribution of slower yet complementary modalities, thereby hindering effective modality fusion and skewing the loss surface toward sharper, less generalizable minima. To address this issue, we propose MBCD, a unified collaborative distillation framework that retains WA's flatness-inducing advantages while overcoming its shortcomings in multi-modal contexts. MBCD begins with adaptive modality dropout in the student model to curb early-stage bias toward dominant modalities. A gradient consistency constraint then aligns learning signals between uni-modal branches and the fused representation, encouraging coordinated and smoother optimization. Finally, a WA-based teacher conducts cross-modal distillation by transferring fused knowledge to each uni-modal branch, which strengthens cross-modal interactions and steer convergence toward flatter solutions. Extensive experiments on MMDG benchmarks show that MBCD consistently outperforms existing methods, achieving superior accuracy and robustness across diverse unseen domains.

Offensive language presents significant challenges on the internet and requires robust moderation. However, the efficacy of such moderation often depends on providing clear and interpretable justifications for each classification. Unfortunately, many existing datasets lack annotated rationales, and most detection models offer limited interpretability and transparency. These limitations hinder the development of trustworthy systems and the implementation of effective content moderation strategies. In this paper, we introduce SomOffXplain, an interpretable framework for detecting offensive language in Somali, which generates human-understandable explanations for its predictions. SomOffXplain performs span-level rationale extraction at both the word and phrase levels, enabling it to highlight text segments that support its predictions. Given that Somali is a low-resource language, we first construct a new benchmark dataset of 10,175 samples, each annotated with human-provided rationales. We evaluate our method against five fine-tuned pre-trained models using Local Interpretable Model-Agnostic Explanations (LIME). Additionally, we adapt four large language models (LLMs) through few-shot and zero-shot prompting to assess their ability to understand and produce rationales in Somali. Our proposed model demonstrates superiority in terms of explainability and predictive accuracy, exhibiting higher plausibility and faithfulness compared to the baselines. Furthermore, our results reveal that half of the state-of-the-art LLMs evaluated fail to generate high-quality rationales that align with human-annotated ground truth rationales, whereas LIME-based methods also prove to be weak explainers for Somali text. We believe our contributions support online safety, help prevent harassment in under-resourced language communities, enhance the trustworthiness of language models, and promote transparency in artificial intelligence systems.

Temporal Knowledge Graph Reasoning (TKGR) plays a vital role in modeling dynamically evolving real-world facts. However, its performance is severely hindered by the long-tail distribution of entities, which induces strong predictive bias against infrequent (tail) entities and significantly undermines the effectiveness of existing methods. To address this challenge, we propose ALSNE, a lightweight and model-agnostic plug-in for adaptive long-tail semantic neighborhood enhancement, designed to alleviate the long-tail issue in TKGR. ALSNE first retrieves semantically similar head entities from the recently interacted neighborhood of each tail entity to provide rich transferable contextual knowledge. This informative knowledge from head entities is then adaptively integrated into the tail entity, facilitating the effective transfer of rich evolutionary patterns that enhance the tail entities’ representation. Extensive experiments conducted on three real-world TKG datasets demonstrate the effectiveness of ALSNE.

Diffusion models excel in generative tasks such as image synthesis and audio processing, yet they are often considered less suitable for regression problems than specialized regression models. In this work, we challenge this view by revealing the potential of diffusion models for controllable speech enhancement through the lens of mutual information. We show that, under specific conditions, diffusion models can outperform traditional regression approaches, and we identify two key obstacles limiting their performance: condition collapse and error accumulation. We address these challenges with three novel, model-agnostic strategies: (1) Condition Prioritization, which emphasizes conditional signals during training; (2) Unfolded Training, which improves robustness against error accumulation; and (3) Attenuated Learning, which stabilizes inference by reducing the impact of intermediate generated samples. Experiments on speech enhancement demonstrate that these strategies consistently enhance performance, even in one-step generation, paving the way for practical and efficient diffusion-based speech enhancement systems.

Self-interpretable graph neural networks (SI-GNNs) provide built-in explanations, yet these explanations are often inconsistent across independently trained models of the same architecture. Our prior work identified redundancy—the inclusion of uninformative features in explanations—as a primary cause of this inconsistency, and demonstrated that averaging explanations from multiple models (Explanation Ensemble, EE) effectively suppresses the high-variance scores of redundant features, thereby improving both explanation consistency and accuracy. Despite its effectiveness, however, EE requires running multiple models to generate consensus explanations, resulting in prohibitive inference cost and incompatibility with standard single-model metrics. In this work, we propose Consensus-driven explanation Distillation (CD), a simple, general, and theoretically grounded framework that leverages ensemble-derived consensus explanations as supervision signals, teaching a single SI-GNN to assign importance scores more responsibly. Experiments across diverse benchmark datasets and SI-GNN frameworks demonstrate that CD retains the benefits of EE while overcoming its limitations, and that combining CD with EE yields additional gains. Our code is publicly available at https://github.com/ICDM-UESTC/ConsensusDistillation.

Social media popularity prediction (SMPP) aims to forecast future engagement of user-generated content and is critical for a wide range of real-world applications. However, existing methods often rely on simple feature fusion or direct use of retrieved neighbors, which can introduce irrelevant signals and overlook cross-modal inconsistencies. To address these limitations, we propose RaMoE, a novel Retrieval-augmented Mixture-of-modal-Experts framework for enhancing SMPP. First, RaMoE employs a field-aware retriever to select historical UGCs that reflect the query’s multimodal context. Second, knowledge-augmented experts align and filter conflicting signals via cross-attention. Finally, a noise-sensitive mixing network adaptively adjusts expert weights to suppress unreliable evidence. Experiments on three real-world datasets demonstrate that RaMoE consistently outperforms state-of-the-art baselines, validating its effectiveness in mitigating retrieval noise and resolving modality conflicts. Our code and data are available at https://github.com/coloreyes/RaMoE.

Network Traffic Anomaly Detection (NTAD) is a critical task in cybersecurity, yet timely and accurate anomaly detection remains challenging. Mamba has emerged as a particularly promising backbone for NTAD due to its linear-time complexity for long-sequence modeling. It further incorporates a dedicated multi-view scanning mechanism to enhance detection precision through complementary contextual cues. However, we identify a previously overlooked structural deficiency in multi-view Mamba scanning for NTAD: redundancy accumulation. Specifically, distinct scanning branches capture substantial view-invariant information, which is repeatedly amplified during multi-view fusion; conversely, view-specific information is diluted or even suppressed, leading to representation homogenization and multi-view degradation. To address this problem, we propose DisenMamba, a novel disentangled multi-view Mamba framework. DisenMamba reformulates multi-view scanning as a two-stage disentangle-then-fuse process that explicitly separates view-invariant and view-specific components prior to fusion. This design prevents the invariant information accumulation while preserving complementary multi-view cues, yielding more discriminative representations for subtle traffic anomalies. Extensive experiments demonstrate the effectiveness of DisenMamba, establishing a new disentangled multi-view Mamba paradigm. Code is available at https://github.com/ikun0124/DisenMamba.

Graph Neural Networks (GNNs) achieve strong predictive performance but offer limited transparency in their decision-making. Self-Interpretable GNNs (SI-GNNs) address this by generating built-in explanations, yet their training objectives are misaligned with evaluation criteria such as faithfulness. This raises two key questions: (i) can faithfulness be explicitly optimized during training, and (ii) does such optimization genuinely improve explanation quality? We show that faithfulness is intrinsically tied to explanation self-consistency and can therefore be optimized directly. Empirical analysis further reveals that self-inconsistency predominantly occurs on unimportant features, linking it to redundancy-driven explanation inconsistency observed in recent work and suggesting untapped potential for improving explanation quality. Building on these insights, we introduce a simple, model-agnostic self-consistency (SC) training strategy. Without changing architectures or pipelines, SC consistently improves explanation quality across multiple dimensions and benchmarks, offering an effective and scalable pathway to more trustworthy GNN explanations.

Deploying multimodal systems in real-world environments often entails handling modality-missing scenarios, where one or more modalities are unavailable. While recent studies address this challenge for the general Multimodal Transformer (MT) architecture via prompt tuning, we identify a fundamental limitation in these methods: the Implicit Modality-Reduction bottleneck. By conditioning prompts solely on the observed modalities, they inadvertently restrict the reasoning scope of MTs to the modality-reduced subspace, cutting off access to the latent information sources of the missing modalities. To overcome this limitation, we propose AOEPT, which pioneers a novel modal-contextualized prompting fashion. Specifically, we introduce lightweight Modal-Contextualized Prompts (MCPs) that distill global modality-wise priors from training data, serving as latent repositories of the information sources for missing modalities. Conditioned on the remaining modalities, these MCPs are instantiated into instance-aware prompts that selectively augment missing-modality information for each sample, thereby restoring the reasoning scope of MTs beyond the observed-modality-only subspace. Experiments across various benchmarks and MT architectures confirm the strong performance of AOEPT, with minimal computational overhead.

Recent work has observed that explanations produced by Self-Interpretable Graph Neural Networks (SI-GNNs) can be self-inconsistent: when the model is reapplied to its own explanatory graph subset, it may produce a different explanation. However, why self-inconsistency arises remains poorly understood. In this work, we first identify re-explanation-induced context perturbation as the direct cause of score variation. We then introduce a latent signal assignment hypothesis to explain why only some edges are sensitive to this perturbation, and analyze how conciseness regularization affects latent signal assignment. Given that self-inconsistent edges do not provide stable evidence for the model's prediction, we propose Self-Denoising (SD), a model-agnostic and training-free post-processing strategy that calibrates explanations with only one additional forward pass. Experiments across representative SI-GNN frameworks, backbone architectures, and benchmark datasets support our hypothesis and show that SD consistently improves explanation quality while adding only about 4–6% computational overhead in practice.

Multi-modal knowledge graph completion (MMKGC) entails predicting missing links in knowledge graphs by integrating both the structural information of triples and the multi-modal features associated with entities. However, existing methods primarily focus on designing various sophisticated multi-modal feature fusion mechanisms while overlooking the impact of both intra- and inter-modal noise (i.e., irrelevant information within modalities and conflicting semantic across modalities). To address these challenges, we propose ROAD, a novel RObust and Adaptive Denoising representation learning framework that mitigates both types of noise to enable more accurate and robust MMKGC. Specifically, to handle intra-modal noise, we introduce an information bottleneck-driven regularization strategy that refines modality-specific entity embeddings by retaining completion-relevant semantics and suppressing irrelevant or noisy information. To alleviate inter-modal noise, we further propose a dual-level consistency-enhanced knowledge integration strategy, which first generates contrastive knowledge-guided hybrid embeddings to ensure semantic consistency across modalities, and then employs confidence-aware decision fusion to adaptively prioritize information from more reliable modalities during prediction. Experiments on three large-scale real-world datasets show that our ROAD consistently outperforms fifteen recent baselines (e.g., compared to the strongest baseline, ROAD achieves improvements of 6.41%, 7.35%, and 2.91% in Hit@1 across all datasets), significantly enhancing the accuracy and robustness of MMKGC.

LiDAR sensors are critical for 3D scene perception in autonomous driving, providing high spatial accuracy but suffering from limited temporal resolution. Point cloud interpolation, which reconstructs intermediate frames from sparsely captured inputs, effectively enhances temporal continuity. However, the challenges posed by non-linear and non-rigid motion in dynamic scenarios remain unresolved by existing methods, which often sacrifice accuracy for efficiency. To solve the above problems, we propose MSTSINet (Multiscale Transformer and SSM Interpolation Network), a novel framework that combines multiscale forward flow prediction with non-linear trajectory compensation for fast and accurate interpolation. Unlike conventional bidirectional flow-based methods, our forward flow-guided approach reduces inference time while improving robustness to complex motion. MSTSINet integrates multiscale motion modeling through a novel hierarchical feature extraction and semantic-aware cross-frame attention, progressively improving forward flow estimation for accurate non-linear trajectory prediction. To further improve motion fidelity and temporal coherence, we introduce a point-wise compensation mechanism during inference and design a novel multiscale loss function that supervises forward flow learning across varying temporal resolutions. Experiments on autonomous driving benchmarks demonstrate that MSTSINet significantly outperforms existing state-of-the-art methods, achieving up to a 26.39% reduction in Chamfer Distance and substantially improving inference speed. These advantages make it well-suited for real-time deployment in autonomous driving scenarios.

Small unmanned aerial vehicles (UAVs) are increasingly used in civilian and industrial applications, making automatic intent understanding essential for intelligent surveillance. Although recent vision-language models can generate captions and answer questions about UAV videos, they often rely on appearance information alone and struggle to reason about motion dynamics, swarm interactions, and the temporal relationships that govern UAV behavior. To address these limitations, we propose a unified multimodal framework that learns an intent-centric representation by jointly integrating complementary visual cues with structured reasoning. Specifically, the framework combines frequency-aware appearance modeling to improve the perception of small and degraded UAVs, blur-guided motion representation to capture motion without relying on optical flow, and spatio-temporal reasoning to model interactions among multiple UAVs before conditioning a large language model for intent-aware caption generation and visual question answering. Rather than treating visual perception and language generation independently, the proposed framework explicitly grounds language understanding in structured intent representations derived from UAV behavior. Experiments on the TD-UAV, NPS-Drones, and Anti-UAV datasets demonstrate consistent improvements over existing methods, highlighting the effectiveness of the proposed framework for UAV intent understanding.

Network Traffic Anomaly Detection (NTAD), particularly under zero-positive settings, is a critical task in cybersecurity. Existing zero-positive NTAD approaches primarily rely on reconstruction-based pipelines. Nevertheless, these methods are susceptible to an identical shortcut issue, where models indiscriminately reconstruct both normal and anomalous inputs, leading to anomaly overgeneralization. To address this limitation, we propose BiPred, the first prediction-based detection paradigm for NTAD. BiPred symmetrically divides each network traffic sample into two segments and reformulates the NTAD task as a bidirectional prediction across these two segments. Unlike reconstruction-based methods, our BiPred not only eliminates shortcut learning by design but also establishes explicit bidirectional contextual dependencies, making it more sensitive to anomalous traffic. Moreover, we develop a novel Residual Scanning Mamba block that encodes multi-view contextual information to support bidirectional prediction. A residual fusion mechanism is proposed for the multi-view scanning Mamba to suppress the accumulation of inter-view redundancy. This design prevents representation degradation and provides multi-view contextual details for prediction. Extensive experiments demonstrate the superiority of our BiPred paradigm, highlighting a new research direction for NTAD. Code is available at https://github.com/ikun0124/BiPred.