Maximum Mean Discrepancy (MMD) is widely used in domain adaptation (DA), yet standard MMD alignment is largely decision-boundary agnostic: reducing cross-domain distributional shift alone does not guarantee a tight target-risk bound. To address this limitation, we propose Decision Boundary-Aware MMD (DB-MMD), a boundary-aware discrepancy that augments MMD with two sample-dependent graphs, namely the Compacting Graph and the Separation Graph, to emphasize boundary-critical samples and enable joint optimization of distribution alignment and decision boundaries. The resulting kernelizable formulation integrates as a plug-in without adversarial training or architectural changes, and admits an efficient generalized-eigenvalue solution. When instantiated in four representative frameworks (JDA, CDDA, DGA-DA, MEDA), DB-MMD yields consistent improvements across eight benchmarks and 49 transfer tasks (up to +9.5 on PIE), achieving competitive or state-of-the-art performance against 50 methods. These results demonstrate that explicitly incorporating boundary awareness into discrepancy learning is an effective and broadly applicable principle for domain adaptation.

Flow matching has recently become a new standard for behavior cloning in robotic manipulation. However, state-of-the-art flow matching policies suffer from a systematic structural mismatch: they rely on a globally fixed isotropic source distribution despite the strongly fragmented and heteroscedastic structure of robotic action spaces. This agnostic initialization forces the model to learn highly entangled vector fields, bottlenecking training efficiency and limiting overall policy performance. To address this limitation, we introduce Latent Action Guided Flow Matching (LAFM), a novel framework that replaces the monolithic Gaussian with an adaptive library of learned prior distributions. By grounding these distributions using a latent action model, LAFM maps current observations to discrete motion primitives, selecting a specialized base distribution that provides an informed, structurally aligned initialization for the denoising process. This dynamic adaptivity naturally accommodates heteroscedasticity in human demonstrations and makes transport trajectories shorter and less entangled. Empirically, LAFM substantially outperforms standard flow matching formulations, increasing task success rates by 23.4

Causal discovery on topological event sequences is crucial for ensuring the reliability of networks. However, existing methods struggle to capture the complex causal relationships arising from concurrent events and lack robustness to incomplete event sequences. To address these issues, we propose a resilient concurrent causal discovery method, termed RCCD, enabling robust learning of causal graphs from topological event sequences. Specifically, we first introduce an influence-aware hyperedge causal attention mechanism, which incorporates event duration into the embedding representation, aggregates concurrent event features via hyperedge causal convolution, and injects network prior knowledge to capture the complex many-to-one causal interactions. Furthermore, we design a masked-based alternating causal optimization framework, which forces the model to recover masked event types based on context through self-supervised mask reconstruction, thereby enhancing the resilience of the predictor to missing data. To validate the effectiveness of our method, we conduct extensive experiments on both simulated and real-world telecommunication network datasets. Experimental results demonstrate that the proposed method significantly outperforms existing state-of-the-art methods in both accuracy and robustness, making it more suitable for real-world telecommunication network environments.

The generalization capabilities of robotic manipulation policies are heavily influenced by the choice of visual representations. Existing approaches typically rely on representations extracted from pre-trained encoders, using two dominant types of features: global features, which summarize an entire image via a single pooled vector, and dense features, which preserve a patch-wise embedding from the final encoder layer. While widely used, both feature types mix task-relevant and irrelevant information, leading to poor generalization under distribution shifts, such as changes in lighting, textures, or the presence of distractors. In this work, we explore an intermediate structured alternative: Slot-Based Object-Centric Representations (SBOCR), which group dense features into a finite set of object-like entities. This representation permits to naturally reduce the noise provided to the robotic manipulation policy while keeping enough information to efficiently perform the task. We benchmark a range of global and dense representations against intermediate slot-based representations, across a suite of simulated and real-world manipulation tasks ranging from simple to complex. We evaluate their generalization under diverse visual conditions, including changes in lighting, texture, and the presence of distractors. Our findings reveal that SBOCR-based policies outperform dense and global representation-based policies in generalization settings, even without task-specific pretraining. These insights suggest that SBOCR is a promising direction for designing visual systems that generalize effectively in dynamic, real-world robotic environments.

Visual foundation models provide strong perceptual features for robotics, but their dense representations lack explicit object-level structure, limiting robustness and contractility in manipulation tasks. We propose STORM (Slot-based Task-aware Object-centric Representation for robotic Manipulation), a lightweight object-centric adaptation module that augments frozen visual foundation models with a small set of semantic-aware slots for robotic manipulation. Rather than retraining large backbones, STORM employs a multi-phase training strategy: object-centric slots are first stabilized through visual–semantic pretraining using language embeddings, then jointly adapted with a downstream manipulation policy. This staged learning prevents degenerate slot formation and preserves semantic consistency while aligning perception with task objectives. Experiments on object discovery benchmarks and simulated manipulation tasks show that STORM improves generalization to visual distractors, and control performance compared to directly using frozen foundation model features or training object-centric representations end-to-end. Our results highlight multi-phase adaptation as an efficient mechanism for transforming generic foundation model features into task-aware object-centric representations for robotic control.

Motivation: Medical image segmentation has become indispensable for clinical diagnosis, preoperative planning, and anatomical investigation. However, existing methods often fail to accurately delineate tiny lesions with multiple targets, complex morphologies, and indistinct boundaries. The core issue lies in their insufficient modeling of fine-grained local spatial relationships and inadequate integration of cross-scale semantic information. These deficiencies prevent the models from capturing subtle structures and edges, ultimately reducing segmentation accuracy.Description: To address the above problems, we propose a medical image segmentation method, DCFANet, based on dynamic context clustering Mamba and context-to-focus attention, which enhances local spatial modeling and cross-scale semantic integration to capture complex boundaries and fine structural details. Specifically, the dynamic context clustering Mamba module mainly comprises the local-global information aggregation module and the dynamic context clustering layer. First, the local-global information aggregation module can adaptively fuse fine-grained textures with high-level semantic features to enhance the representation of lesion boundaries and microscopic structures. Second, the dynamic context clustering layer dynamically partitions image tokens into subregions and performs adaptive clustering within each subregion. This enables region-specific feature interactions that preserve local spatial consistency and capture subtle structural variations. Finally, the context-to-focus attention module can achieve precise alignment and semantic fusion of features between the encoder and decoder.Results and conclusion: Extensive experimental evaluations on five diverse medical image segmentation datasets demonstrate that DCFANet consistently achieves superior performance compared to existing methods, validating its wide applicability and exceptional generalization ability.

Cross-dataset Human Activity Recognition (HAR) suffers from limited model generalization, hindering its practical deployment. While prior research has predominantly focused on designing novel model architectures, the crucial role of training data composition is often overlooked. To address this gap, inspired by the success of data mixture optimization in Large Language Models (LLMs), we introduce this strategy to optimize the composition of multi-source training data for HAR models. This approach facilitates the creation of balanced and effective training datasets, thereby enhancing the models’ utility across diverse conditions. To tailor this strategy to continuous, multi-channel Inertial Measurement Unit (IMU) data, we propose HAR-DoReMi, a framework consisting of two complementary parts: a domain reweighting scheme and a pre-processing step for sensor alignment. The domain reweighting scheme includes a Smooth-Stable Strategy using a pre-computed global baseline to improve training stability, a Conditional Batch Score (CBS) for more accurate domain difficulty assessment, and a composite Mean Squared Error (MSE) and Soft-Dynamic Time Warping (Soft-DTW) loss to capture both signal fidelity and temporal dynamics. In addition, the pre-processing step employs the Mahony fusion algorithm to reduce sensor orientation heterogeneity, thereby complementing the domain reweighting scheme in cross-dataset HAR. Extensive experiments across multiple cross-dataset transfer tasks show that the full HAR-DoReMi framework improves average accuracy by 10.5% over baseline approaches while using about 30% to 50% of the training data, improving both robustness under distribution shift and data efficiency. Code will be available at https://github.com/ABan147/har-doremi.

Imitation learning has enabled remarkable progress in robotic manipulation, especially with diffusion and flow-based policies that generate complex visuomotor behaviors directly from demonstrations. Yet, despite their strong performance, these policies often fail to generalize across tasks and environments. A key reason is that existing policies tend to imitate superficial action correlations rather than the underlying intent. Inspired by the compositional structure of human behaviors, we propose PriGo, a primitive-guided test-time adaptive framework for robust robotic manipulation. PriGo introduces PANet, a lightweight primitive prediction module that infers primitive distributions directly from observations. We further propose a differentiable primitive guidance mechanism that refines generated actions during inference, steering trajectories toward semantically consistent behaviors. Unlike prior primitive-conditioned approaches, PriGo operates entirely at test time and can be seamlessly integrated into pretrained diffusion and flow policies without retraining. Extensive experiments on LIBERO, CALVIN, SIMPLER, and real-world robotic tasks demonstrate that PriGo consistently improves robustness, long-horizon execution, and generalization ability across both diffusion and flow-based policies.

Robotic manipulation policies rely on pre-trained vision models that give either a global scene embedding or a dense patch grid. Both mix task-relevant and task-irrelevant features. Object-centric slot representations are a structured alternative: they group features into a few per-object slots. We test what this structure buys on ManiSkill3 PickCube-v1, with a frozen encoder and a held-out-seed evaluation. Holding the policy, goal token, rendering, and calibration fixed and changing only the encoder, a frozen object-centric SPOT representation (DINO ViT-B/16 + Slot Attention) reaches 55.0±2.9

Variational autoencoders (VAEs) have shown great promise in recommender systems due to their advantage of handling implicit feedback. However, existing VAE-based methods still rely on (i) static, data-independent Gaussian priors that fail to reflect actual user-item interaction dynamics, and (ii) Euclidean embeddings that distort the power-law structure of interaction patterns. To address these limitations, we propose Hyperbolic Adversarial Variational Embedding (HAVE), a unified framework that combines adversarial variational inference, data-driven prior adaptation, and non-Euclidean representation learning. First, we introduce an adversarial variational inference paradigm that matches the encoder's posterior to a richer, learnable target distribution, thereby enhancing flexibility and capturing nuanced interest semantics. Building on this, we design an Adaptively Variational Prior (AdaPrior) which fuses user activity patterns and item popularity trends into behavior-aware priors, mitigating posterior collapse. Finally, HAVE embeds user behaviors into hyperbolic space, exploiting its hierarchical structure to naturally encode both a small core of highly active users and a vast periphery of casual users without distortion. Extensive experiments on multiple real-world datasets demonstrate that HAVE not only outperforms state-of-the-art baselines in recommendation accuracy but also yields latent representations that faithfully preserve hierarchical interaction structures.

Active perception is a fundamental skill that enables us humans to deal with uncertainty in our inherently partially observable environment. For senses such as touch, where the information is sparse and local, active perception becomes crucial. In recent years, active perception has emerged as an important research domain in robotics. However, current methods are often bound to specific tasks or make strong assumptions, which limit their generality. To address this gap, this work introduces APPLE (Active Perception Policy Learning) - a novel framework that leverages reinforcement learning (RL) to address a range of different active perception problems. APPLE jointly trains a transformer-based perception module and decision-making policy with a unified optimization objective, learning how to actively gather information. By design, APPLE is not limited to a specific task and can, in principle, be applied to a wide range of active perception problems. We evaluate two variants of APPLE across different tasks, including tactile exploration problems from the Tactile MNIST benchmark. Experiments demonstrate the efficacy of APPLE, achieving high accuracies on both regression and classification tasks. These findings underscore the potential of APPLE as a versatile and general framework for advancing active perception in robotics.

Online planning in continuous partially observable Markov decision processes (POMDPs) using ω-regular specifications requires handling continuous belief dynamics within the finite symbolic memory in order to track temporal progress. Existing methods based on either direct search in belief space or predefined discrete abstractions suffer from drawbacks, e.g., lack of symbolic memory for long-horizon logical progress or difficult to certify from noisy online beliefs. As such, obtaining reliable symbolic states online from continuous observations remains a challenge. To address this issue, we introduce the Revealed Belief Automaton (REBA), an event-driven framework that advances the research from global belief-space discretization to a fundamental new way of thinking, namely online certification of revelation events. Specifically, we propose an online revelation method that, through information-theoretic gates, can dynamically analyse and establish belief abstraction from the continuous belief space by discovering reliable anchors among noisy beliefs. We then develop an incremental topology adaptation mechanism over the certified anchors to realise the online finite Belief Automaton. By combining with the ω-regular specification, REBA is able to support formal parity policy synthesis without a predefined discrete abstraction, which in turn can guide the Monte Carlo Tree Search process to perform online search beyond its local horizon. In addition, we design an error decomposition analysis which can assess the effectiveness and reliability of this discrete guidance for the underlying continuous POMDP. Empirical evaluations in patrolling and navigation scenarios show that REBA matches or exceeds all evaluated baselines, with primary metric gains of +17.0% to +47.4% over state-of-the-art approaches.

Deploying object-centric models for real-world scene understanding typically requires complex pipelines to achieve both robust scene decomposition and high-fidelity generation. Recent diffusion-based approaches have improved visual quality, but they almost universally rely on heavy, pretrained generative priors (e.g., Stable Diffusion) and external VAE latent spaces. In this paper, we propose Slot-RAE, a much simpler, fully integrated framework that operates directly within the continuous semantic feature space of visual foundation models (e.g., DINOv3). Slot-RAE employs a feature-space diffusion process using a Diffusion Transformer (DiT) decoder and a Representation Alignment (REPA) head. Unlike existing diffusion-based objectcentric methods that rely heavily on subsidized text-toimage priors, the generative core of Slot-RAE (Slot Attention and the DiT) is trained from scratch within the frozen VFM feature space. This eliminates the need for VAE bottlenecks and task-agnostic generative pre-training. Experiments on the COCO dataset demonstrate that despite its architectural simplicity, Slot-RAE achieves state-of-the-art results. It delivers comparable unsupervised object discovery, higher-fidelity image reconstruction, and robust zero-shot compositionality, all while being significantly faster and more computationally efficient than existing object-centric latent diffusion models.

Recently robot-assisted deep brain stimulation has been widely applied to alleviate symptoms of Parkinson patients. Considering that the surgical outcome is highly associated with preoperative planning, while the present electrode implantation planning is manually performed on 2D images, which is time-consuming and highly dependent on neurosurgeon’s experience. To tackle above issues, we present a novel electrode insertion planning pipeline for precise robotic surgery. First, a data and knowledge co-driven framework is proposed to effectively associate the patients’ brain and “secure implantation channel”. Then, an automatic method is introduced to determine the locations of the stimulation targets and structures at risk during the surgery. Besides, the corresponding signed distance fields are generated for measuring the closest distances from electrode to structures at risk. In addition, we proposed a multi-constraints based planning method in data and knowledge co-defined region to search the optimal electrode insertion path while avoiding cerebral obstacles. We evaluated our method on 10 PD patients undergoing DBS treatment. Our pipeline achieved average obstacle-avoidance distances of 2.15 mm for sulci and 5.68 mm for ventricles, outperforming manual planning by surgeons. Consequently, our approach can produce clinically significant plans with precise stimulation targets and secure insertion trajectories for neurosurgical robotics, demonstrating strong potential for clinical application.

Multimodal classification is a pivotal research field in artificial intelligence, aiming to improve model performance by combining complementary information across different modalities, such as images and text. However, previous approaches have focused on either intra-modal relationships while overlooking the potential of cross-modal complementarity or inter-modal interactions while neglecting information within specific modalities. To address these limitations, this paper proposes a multimodal transformer fusion framework based on intra-modal local and inter-modal global attention (ILIGA) effectively integrating information both within and across modalities. First, an intra-modal local attention fusion module is employed to capture local information and strengthen feature representation within each modality. Then, an inter-modal global outer product fusion module is applied to model global complementary relationships across modalities. Extensive experiments conducted on ISIC2019 and UPMC-Food-101 public datasets demonstrate that the proposed model has achieved superior performance in image–text multimodal classification tasks, with accuracies of 80.42% and 92.05%, respectively. Furthermore, ablation studies are performed to verify the effectiveness of each component in our model.

Proactively predicting a users next utterance in human-machine dialogue can streamline interaction and improve user experience. Existing commercial API-based solutions are subject to privacy concerns while deploying general-purpose LLMs locally remains computationally expensive. As such, training a compact, task-specific LLM provides a practical alternative. Although user simulator methods can predict a user's next utterance, they mainly imitate their speaking style rather than advancing the dialogue. Preference data synthesis has been investigated to generate data for proactive next utterance prediction and help align LLMs with user preferences. Yet existing methods lack the ability to explicitly model the intent reasoning that leads to the user's next utterance and to define and synthesize preference and non-preference reasoning processes for predicting the user's next utterance.To address these challenges, we propose ProUtt, an LLM-driven preference data synthesis method for proactive next utterance prediction. ProUtt converts dialogue history into an intent tree and explicitly models intent reasoning trajectories by predicting the next plausible path from both exploitation and exploration perspectives. It then constructs preference and non-preference reasoning processes by perturbing or revising intent tree paths at different future turns. Extensive evaluations using LLM-as-a-judge and human judgments demonstrate that ProUtt consistently outperforms existing data synthesis methods, user simulators, and commercial LLM APIs across four benchmark datasets. We release both the code and the synthesized datasets to facilitate future research.

Medical image segmentation techniques play a vital role in clinical diagnosis, surgical planning, and postoperative evaluation. However, medical images typically exhibit highly complex nonlinear spatial correlations, and their irregular geometric morphologies and indistinct boundary features make accurate segmentation highly challenging. The methods based on the convolutional neural network, Transformer, and Mamba struggle to effectively capture high-order dependencies in complex anatomical structures. Although the method based on hypergraphs can alleviate this problem, most of the existing methods are limited to static modeling, which not only ignores the complementarity and topological consistency among hypergraphs, but also lacks explicit modeling of the dynamic dependency relationships among hyperedges. To this end, this paper proposes a novel medical image segmentation method based on dynamic hypergraph modeling and Mamba, termed HG-Mamba. Specifically, we propose the multi-view hypergraph aggregation module to simultaneously capture the cross-structure high-order dependencies at the global level and the fine-grained high-order correlations within local neighborhoods. Secondly, we propose the dynamic hypergraph convolution. It enhances the propagation of group-level contextual information by explicitly modeling high-order contextual interactions among hyperedges. Furthermore, we introduce the multi-scale dual-domain aggregation module. It jointly captures local details and global contour features through dual constraints in both spatial and frequency domains. To evaluate the segmentation performance of HG-Mamba, we conduct experiments on four datasets. The results demonstrate that HG-Mamba outperforms current state-of-the-art methods in both segmentation accuracy and generalization ability.

Sodium niobate-based dielectric energy storage materials, as key components in capacitors, have the advantages such as low relative density, lead-free, low cost, and excellent energy storage density, and can meet the important requirements of electronic components for miniaturization, harmlessness, integration and light weight. Therefore, they have received extensive attention from the scientific community in recent years. In this work, by introducing both Bi(Mg0.5Sn0.5)O-3 and (Bi0.5Na0.5)(0.7) Sr(0.3)TiO(3 )components into NaNbO3 ceramics, a conventional solid-phase sintering method is used to prepare (1-x)[0.93NaNbO(3)-0.07Bi(Mg0.5Sn0.5)O-3]- x (Bi-0.5 Na-0.5)(0.7)Sr0.3TiO3(Abbreviated as (1-x)(NN-BMS)-xBNST, 0 <= x <= 0.3) relaxation ferroelectric ceramics, and the ceramics are characterized by using X-ray diffraction, scanning electron microscopy, UV spectroscopy and Raman spectroscopy so as to study the effects of (Bi0.5Na0.5)(0.7)Sr0.3TiO3 doping on the physical phase composition, microstructure, and electrical properties of NaNbO(3 )ceramics, such as dielectric and energy storage. The (1-x)(NN-BMS)-xBNST ceramics exhibit a single perovskite structure, with cell volume a first increasing and then decreasing. The coexistence of Pbma and Pnma phases (1-x)(NN-BMS)-xBNST ceramics exhibits a dense microstructure and clear grain boundaries at an optimal sintering temperature. The average grain size first increases to 4.73 mu m, then decreases to 2.17 mu m, and finally increases to 3.06 mu m. A smaller grain size and a larger bandgap width are beneficial for improving the breakdown strength. The 0.75(NN-BMS)0.25BNST ceramic shows the excellent dielectric temperature stability (25-160 degrees C, Delta epsilon/epsilon 25 degrees C <= +/- 15%) and dielectric frequency stability, which can meet the EIAZ8U standard and hence work in a special environment (high temperature and high frequency). Meanwhile, 0.75(NN-BMS)-0.25BNST ceramic exhibits excellent energy storage performance at high field strength (390 kV/cm): recoverable energy density W rec = 2.73 J/cm(3), energy storage efficiency eta = 82.6%, and high temperature stability in a temperature range of 20-100 degrees C. The research results indicate that 0.75(NN-BMS)-0.25BNST ceramics have broad prospects of applications in lead-free dielectric energy storage capacitors.

3D face recognition (3DFR) algorithms have advanced significantly in the past two decades by leveraging facial geometric information, but they mostly focus on high-quality 3D face scans, thus limiting their practicality in real-world scenarios. Recently, with the development of affordable consumer-level depth cameras, the focus has shifted towards low-quality 3D face scans. In this paper, we propose a method for low-quality 3DFR. On one hand, our approach employs the normalizing flow to model an adaptive-form distribution for any given 3D face scan. This adaptive distributional representation learning strategy allows for more robust representations of low-quality 3D face scans (which may be caused by the scan noises, pose or occlusion variations, etc.). On the other hand, we introduce an adaptive sample weighting strategy to adjust the importance of each training sample by measuring both the difficulty of being recognized and the data quality. This adaptive sample weighting strategy can further enhance the robustness of the deep model and meanwhile improve its performance on low-quality 3DFR. Through comprehensive experiments, we demonstrate that our method can significantly improve the performance of low-quality 3DFR. For example, our method achieves competitive results on both the IIIT-D database and the Lock3DFace datasets, underscoring its effectiveness in addressing the challenges associated with low-quality 3D faces.

With the rapid proliferation of the Internet of Things (IoT), robust trust management has become imperative to ensure security in these IoT systems. Prior machine learning approaches to IoT trust management have exhibited suboptimal performance, failing to capture the dynamic behaviour and complex discriminative features of IoT devices. To address these challenges, we design an evocative trust management scheme for user’s authentication based on Dempster–Shafer’s evidence theory, which can persuade the normal activities of IoT device systems. We establish a set of discriminating features to predict the trustworthiness of a network node by assessing its observed behaviours. These behaviours being assessed encompass several characteristics such as throughput, delay, jitter and network latency. As such, nodes that demonstrate elevated data transmission rates and have anomalous traffic patterns could potentially be categorised as untrustworthy. In addition, the existence of persistently high latency will impede trust prediction algorithms, and this will consequently alter the overall behaviour of the node, ultimately affecting the trustworthiness of the entire network. We also design a framework for the fusion of evidence based on the belief degree and reputation-based evidence to avoid misclassification resulting in evidence conflicting. The resultant fusion outcomes are transformed into category labels which serve as the prediction outcome of the multi-classifier ensemble scheme. We evaluate our proposed scheme with and without the best discriminative features on performance metrics including accuracy, precision, recall, F1-score, detection rate and false alarm rate. Comprehensive experiments on a transformed UNSW-NB15 dataset demonstrate the better performance of our proposed framework, especially in the application of evidence conflicting.