Multi-Modal Anomaly Detection (MMAD) detects rare abnormal events from heterogeneous data sources and is increasingly used in safety- and reliability-critical applications such as industrial inspection and cybersecurity. Yet the literature is fragmented across domains and modality combinations, and existing surveys usually group methods by architecture rather than by how abnormality is defined and separated in multi-modal settings. We survey MMAD from an assumption-driven perspective. We formalize the problem, identify five intrinsic characteristics underlying its core challenges, and organize prior work into two complementary paradigms. The first, normality-assumption methods, models regularity via representation learning, cross-modal alignment, and knowledge enhancement. The second, anomaly-assumption methods, sharpens decision boundaries through coarse-grained, structural, and semantic anomaly injection. We also investigate how foundation models are reshaping MMAD through scalable pretraining, flexible cross-modal transfer, and emerging reasoning capabilities. Finally, we compile representative benchmarks and evaluation protocols across domains and highlight open problems and future directions for robust, adaptive, and interpretable MMAD systems.

Question answering over knowledge graphs (QA-KG) is a challenging task due to the complexity of natural language questions. Competitive methods proposed for the task utilize reinforcement learning approaches to perform multi-hop reasoning over questions. These methods often use weak supervision signals from answer entities and manually set rewards to guide the model during the reasoning process. However, these methods open the risk of spurious trajectories that incidentally lead to answer entities, which eventually leads to negative supervision. Unfortunately, these trajectories are not reflected in the existing evaluation indicators to validate the reasoning process. As a solution, we propose a Progressive Multi-hop Reasoning (PMR) model for QA-KG. The model is composed of an encoder, selector, linker and checker, which learns to hop progressively in the reasoning process by leveraging strong supervision signals from SPARQL parsing results. Extensive experiments show that our model achieves state-of-the-art results on two mainstream datasets. More importantly, our results far exceeds methods based on reinforcement learning when matching the intermediate relationship path.

The increasing volume of time series signals and the scarcity of labels make time series anomaly detection (TSAD) a natural fit for self-supervised deep learning. However, existing normality-based approaches face two key limitations: relying on a single assumption often fails to capture the whole normality, leading to biased representations; and they typically presume clean training data, which is unrealistic in practice and undermines model robustness. In this article, we propose RoCA, a unified and robust anomaly detection (AD) framework that simultaneously addresses assumption incompleteness and data contamination. The key insight is that normal samples tend to satisfy multiple normality assumptions, whereas anomalous or contaminated samples should violate at least one. RoCA employs a composite loss function consisting of a multinormality alignment term, a dynamic anomaly-aware term, and a variance term to maintain training stability. This design enables RoCA to dynamically discover and push away latent anomalies during training to refine the boundary, eliminating the dependence on precisely labeled, high-purity training data. Extensive experiments on both univariate and multivariate time series datasets demonstrate that RoCA consistently outperforms state-of-the-art methods, achieving up to 7.3% improvement under real-world contamination. Our theoretical analysis further reveals the intrinsic synergy between contrastive learning (CL) and one-class classification (OC) under the RoCA framework. The source code is available at the anonymous repository https://github.com/ruiking04/RoCA

Dialogue policy, a critical component in multi-domain task-oriented dialogue systems, decides the dialogue acts according to the received dialogue state. We introduce zero-shot reinforcement learning for dialogue policy learning, which aims to learn dialogue policies capable of generalizing to unseen domains without further training. This setup brings forward two challenges: 1) the representation of unseen actions states, and 2) zero-shot generalization to unseen domains. For the first issue, we propose Unified Representation (UR), an ontology-agnostic representation, which effectively infers representations in unseen domains by capturing the underlying semantic relations between unseen actions and states and seen ones. To tackle the second issue, we propose Q-Values Perturbation (QVP), a family of exploration strategies that can be applied either during training or testing. Experiments on MultiWOZ, suggest that UR, QVP, and an integrated framework combining the two are all effective.

Graph Neural Networks (GNNs) have found widespread application in malware detection tasks in recent years, aiming to uncover the malicious nature of target processes by aggregating neighborhood information via different relations. However, current GNN-based malware detection models are primarily designed for static graphs or fixed-window dynamic graphs; they often overlook the correlation between distinct continuous behaviors and specific attacks (e.g., code injection), leading to a disruption in the attack's continuity and incurring substantial computing overhead. This paper introduces a novel model dubbed DynMD, designed to enhance the efficiency of malware detection on streaming behavioral data via energy-based dynamic graph representation learning. Concretely, DynMD investigates an energy function to adaptively partition windows, facilitating the construction of a lossless malware dynamic heterogeneous graph. More importantly, DynMD proposes a unique time-aware dynamic graph learning method to capture energy-concordant dynamic neighborhoods. This approach involves both intra-graph and inter-graph message propagation, promoting dynamic graph learning while avoiding time-consuming and repetitive computations. Experimental results on three real-world malware datasets show that DynMD improves detection accuracy by up to 3.99% and achieves detection 3.81x to 5.33x faster than the MG-DVD framework, identifying malware within an average of 40 seconds of execution.

Time series anomaly detection is essential for applications such as finance and healthcare. However, existing methods often struggle to capture complex multi-scale dependencies or are trapped in capturing low-level sample characteristics. To address the issue, we propose MOC, a novel multi-scale one-class anomaly detection framework that integrates Mamba and CNN to model global dependencies and local dynamics jointly. In addition, MOC introduces a one-class to compact normal patterns into a hypersphere, by a hinge loss to prevent hypersphere collapse, improving robustness against noise. Extensive experiments on three public datasets show that MOC significantly outperforms state-of-the-art baselines, demonstrating its effectiveness for robust and accurate performance.

Redundant manipulators, with their higher Degrees of Freedom (DoFs), provide improved kinematic capability and versatility, making them well-suited to applications such as manufacturing, surgical robotics, and human-robot collaboration. However, motion planning for redundant manipulators remains challenging due to the increased dimensionality of the configuration space and the complexity of dynamic environments. Traditional motion planning algorithms often struggle to scale to such high-dimensional problems, while deep learning-based planners can be unstable or inefficient on complex tasks. This paper presents RobotDiffuse, a diffusion model-based motion planning framework for redundant manipulators. RobotDiffuse incorporates physical constraints through a point cloud encoder, and replaces the U-Net backbone with an encoder-only transformer to better capture temporal dependencies, producing smoother and more coherent motion plans. We evaluate the proposed method in a challenging simulator and release a new dataset, Robot-obstacles-panda (ROP), containing 35M robot poses and 0.14M obstacle-avoidance scenarios. RobotDiffuse achieves the highest overall score in our experiments, demonstrating both its effectiveness and the potential of diffusion models for motion planning. The dataset is available at https://github.com/ACRoboT-buaa/RobotDiffuse .

The Satellite-Terrestrial Integrated Network (STIN) technology based on low-earth orbit (LEO) satellites provides a seamless network with low latency and high reliability to global users. The Space-Terrestrial Collaborative Computing (STCC) paradigm based on STIN has become a promising solution for ubiquitous computing. Due to the mobility of satellites and the vulnerability of inter-satellite and satellite-terrestrial network connections, the research on resource scheduling strategy, configuration, and deployment of STCC services is more complex than that of ground cloud services. The lack of simulators further restricts the development of STCC research. In order to solve this problem, we propose an STCC simulation tool for the satellite-terrestrial hybrid cloud in this paper. The tool can simulate the hybrid network models and support the simulation of the flow computing model. By evaluating the performance of the task offloading policy, we demonstrate the effectiveness of the simulation tool.

With the rapid scaling of large language model (LLM) training clusters, GPU errors frequently occur and disrupt the training process. While traditional error recovery methods, such as periodic checkpointing, are effective, they incur substantial overhead in both daily operations and recovery processes. Just-in-time checkpointing, a representative alternative, reduces this overhead by eliminating the periodic checkpoint saving procedure and optimizing the recovery workflow. However, its complex GPU context decoupling mechanism and global reinitialization of communication backend remain resource-intensive and slow. In this paper, we present Mnemosyne, a lightweight and fast error recovery framework for LLM training. To minimize both daily and recovery overhead, Mnemosyne optimizes the GPU context decoupling component with shared-memory-based IPC and index-based handle mapping, and designs a flexible collective communication library that dynamically adjusts links of built communicators without requiring reinitialization. Preliminary experiments on our open-source prototype demonstrate that, compared to the state of the art, Mnemosyne reduces daily overhead by up to 58.8% and communication rebuilding time by up to 91.3%.

The accumulation of time-series signals and the absence of labels make time-series Anomaly Detection (AD) a self-supervised task of deep learning. Methods based on normality assumptions face the following three limitations: (1) A single assumption could hardly characterize the whole normality or lead to some deviation. (2) Some assumptions may go against the principle of AD. (3) Their basic assumption is that the training data is uncontaminated (free of anomalies), which is unrealistic in practice, leading to a decline in robustness. This paper proposes a novel robust approach, RoCA, which is the first to address all of the above three challenges, as far as we are aware. It fuses the separated assumptions of one-class classification and contrastive learning in a single training process to characterize a more complete so-called normality. Additionally, it monitors the training data and computes a carefully designed anomaly score throughout the training process. This score helps identify latent anomalies, which are then used to define the classification boundary, inspired by the concept of outlier exposure. The performance on AIOps datasets improved by 6 (COCA). On two large and high-dimensional multivariate datasets, the performance increased by 5 performance on both univariate and multivariate datasets. The source code is available at https://github.com/ruiking04/RoCA.

Time-series anomaly detection is crucial in AIOps for maintaining large-scale service reliability. In production, streams of Key Performance Indicators (KPI) are high-dimensional, non-stationary, and affected by noise, deployment changes, and latent anomalies, making real failures hard to distinguish from benign variation. Most existing methods assume either normality (learning from "normal" history) or rely on injected anomalies for training. Yet injected patterns often misalign with real failure modes, skewing decision boundaries – aka. Anomaly Shift. We propose CAPMix, a controllable anomaly augmentation framework with prior-guided injection for realistic temporal behaviors. CAPMix combines label revision and dual-space mixup to enhance robustness under contaminated and mixed data. CAPMix consistently outperforms state-of-the-art methods on public AIOps and time-series benchmarks. It has been deployed in Kuaishou's large-scale production system, reducing false alarms and improving monitoring reliability. A real-world dataset is also released to enrich the research on robust KPI anomaly detection.

Pareto Front Learning (PFL) has been one of the effective means to resolve multi-objective optimization problems through exploring all optimal solutions to learn the entire Pareto front. Pareto Hypernetwork (PHN) is a new promising way to generate the sequence of Pareto-optimal solutions that can be further used as potential solutions to constitute the Pareto front. However, the existing PHN-based approaches suffer from two performance issues: They take as inputs human-crafted preference vector or chunk embedding, rather than the input data samples, and thus vulnerable to data distribution shifts. Such approaches cannot optimize all potential solutions when forming the Pareto front, as they merely optimize the loss pertaining to one single input at a time of optimization round. To improve the quality of the Pareto front, we propose IOP, a novel Idempotent-like Optimization method to learn the entire Pareto front accurately and enhance Hypernetwork's adaptability to distribution shifts. In particular, IOP performs idempotent-like optimization by exploiting manifold space mapping, so that the target networks generated by the optimized Hypernetwork can effectively handle samples with similar distributions of the input samples, without the pre-defined human-crafted inputs. IOP maximizes the Hypervolume indicator that is composed of all potential solutions at a higher level. Experimental results demonstrate that IOP outperforms the state-of-the-art methods by 4.7% on average in producing the Pareto front and has a 10.5% improvement in adaptability.

Time series anomaly detection is challenging due to the rarity of anomalies and the complexity and diversity of normal patterns. Most existing methods rely on a single hypothesis and learn feature patterns from a limited dataset, which restricts their generalization capabilities. At the same time, time series foundation models have shown promising results across multiple tasks due to their strong generalization capabilities. However, time series foundation models are less likely to achieve better performance in complex anomaly detection tasks. To address this issue, this paper introduces FOCA, a novel foundation-model-based one-class anomaly detection approach. This method preserves the generalization ability of the foundation model to capture normal variation patterns and provides a comprehensive feature space for one-class classification. It constrains normal features within a sufficiently small hypersphere to construct a decision boundary for detecting abnormal data. Furthermore, it is observed in practice that the introduction of fine-tuning techniques can further improve the performance of the method. Extensive experiments on two standard benchmark datasets demonstrate that our method outperforms the state-of-the-art approaches.

The growing adoption of industrial robots has spurred interest in combining Industrial IoT and cloud computing technologies to create unified management platforms. Traditional systems focus mainly on analyzing robot-status data to infer maintenance cycles and monitor operational health, yet often overlook the complex process optimization tasks that many robots must execute. In addition, the heterogeneity among industrial robot operating systems further complicates the management and execution of process optimization. To address these challenges, we propose CIRPO-a Cloud-edge collaborative platform for Industrial Robots Process Optimization, a framework designed to unify both the execution and optimization of robotic processes across cloud and edge tiers. It provides a unified command interface by abstracting vendor-specific controller differences, and implements process dispatching and execution through a workflow-oriented approach, thereby supporting diverse industrial operations. We validate the framework by integrating two real-world applications-palletizing and polishing-demonstrating CIRPO's ability to deliver closed-loop, cloud-based process optimization for heterogeneous robot fleets. The CIRPO source code, helm charts, and anonymised experiment traces will be release on github upon acceptance.

Effective anomaly detection in time series is pivotal for modern industrial applications and financial systems. Due to the scarcity of anomaly labels and the high cost of manual labeling, reconstruction-based unsupervised approaches have garnered considerable attention. However, accurate anomaly detection remains an unsettled challenge, since the optimization objectives of reconstruction-based methods merely rely on point-by-point distance measures, ignoring the potential structural characteristics of time series and thus failing to tackle complex pattern-wise anomalies. In this paper, we propose StrAD, a novel structure-enhanced anomaly detection approach to enrich the optimization objective by incorporating structural information hidden in the time series and steering the data reconstruction procedure to better capture such structural features. StrAD accommodates the trend, seasonality, and shape in the optimization objective of the reconstruction model to learn latent structural characteristics and capture the intrinsic pattern variation of time series. The proposed structure-aware optimization objective mechanism can assure the alignment between the original data and the reconstructed data in terms of structural features, thereby keeping consistency in global fluctuation and local characteristics. The mechanism is pluggable and applicable to any reconstruction-based methods, enhancing the model sensitivity to both point-wise anomalies and pattern-wise anomalies. Experimental results show that StrAD improves the performance of state-of-the-art reconstruction-based models across five real-world anomaly detection datasets.

Accurately identifying network attacks is crucial for maintaining network security. However, these attacks are often hide within massive volumes of network traffic, posing significant challenges for traditional detection methods. Supervised learning approaches require substantial labeled data and struggle to adapt to unknown attack types, while unsupervised methods face difficulties in accurately pinpointing specific attack categories. To address these limitations, we propose a novel fewshot learning model for network flow attack classification based on graph contrastive learning. Our model leverages contrastive learning to enhance feature representation and generalization capabilities, enabling high-accuracy attack detection even with limited training data. Specifically, we first construct a multi- graph representation of network traffic and segment the data into snapshots. Then, we perform graph data augmentation within each snapshot to generate augmented sample pairs, which are used to pre-train the model via contrastive learning. Finally, we fine-tune the model parameters to achieve multi-class attack classification, leveraging the learned feature representations to identify various attack types, even those unseen during training. Experimental results demonstrate that our model exhibits excellent generalization ability and achieves high attack detection performance, even with limited training data.

As the application of Industrial Robots (IRs) scales and related participants increase, the demands for intelligent Operation and Maintenance (O M) and multi-tenant collaboration rise. Traditional methods could no longer cover the requirements, while the Industrial Internet of Things (IIoT) has been considered a promising solution. However, there’s a lack of IIoT platforms dedicated to IR O M, including IR maintenance, process optimization, and knowledge sharing. In this context, this paper puts forward the multi-tenant-oriented ACbot platform, which attempts to provide the first holistic IIoT-based solution for O M of IRs. Based on an information model designed for the IR field, ACbot has implemented an application architecture with resource and microservice management across the cloud and multiple edges. On this basis, we develop four vital applications including real-time monitoring, health management, process optimization, and knowledge graph. We have deployed the ACbot platform in real-world scenarios that contain various participants, types of IRs, and processes. To date, ACbot has been accessed by 10 organizations and managed 60 industrial robots, demonstrating that the platform fulfills our expectations. Furthermore, the application results also showcase its robustness, versatility, and adaptability for developing and hosting intelligent robot applications.

Detecting time-series anomalies is extremely intricate due to the rarity of anomalies and imbalanced sample categories, which often result in costly and challenging anomaly labeling. Most of the existing approaches largely depend on assumptions of normality, overlooking labeled abnormal samples. While anomaly assumptions based methods can incorporate prior knowledge of anomalies for data augmentation in training classifiers, the adopted random or coarse-grained augmentation approaches solely focus on pointwise anomalies and lack cutting-edge domain knowledge, making them less likely to achieve better performance. This paper introduces CutAddPaste, a novel anomaly assumption-based approach for detecting time-series anomalies. It primarily employs a data augmentation strategy to generate pseudo anomalies, by exploiting prior knowledge of anomalies as much as possible. At the core of CutAddPaste is cutting patches from random positions in temporal subsequence samples, adding linear trend terms, and pasting them into other samples, so that it can well approximate a variety of anomalies, including point and pattern anomalies. Experiments on standard benchmark datasets demonstrate that our method outperforms the state-of-the-art approaches.

Current methods for Knowledge-Based Question Answering (KBQA) usually rely on complex training techniques and model frameworks, leading to many limitations in practical applications. Recently, the emergence of In-Context Learning (ICL) capabilities in Large Language Models (LLMs) provides a simple and training-free semantic parsing paradigm for KBQA: Given a small number of questions and their labeled logical forms as demo examples, LLMs can understand the task intent and generate the logic form for a new question. However, current powerful LLMs have little exposure to logic forms during pre-training, resulting in a high format error rate. To solve this problem, we propose a code-style in-context learning method for KBQA, which converts the generation process of unfamiliar logical form into the more familiar code generation process for LLMs. Experimental results on three mainstream datasets show that our method dramatically mitigated the formatting error problem in generating logic forms while realizing a new SOTA on WebQSP, GrailQA, and GraphQ under the few-shot setting. The code and supplementary files are released at https://github.com/Arthurizijar/KB-Coder .