Deep research benchmarks require expert-level tasks and reliable evaluation grounded in task-specific knowledge. Existing benchmarks rely heavily on expert authoring or pre-existing human-authored materials, while fully automatic construction struggles to ensure consistent and traceable verification. To address this gap, we introduce a verifiable benchmark of 500 deep research tasks spanning 31 topics and 10 major categories, with three query forms designed to probe complementary capabilities required for deep research. The benchmark is constructed automatically using an iterative Explorer-Formalizer-Challenger pipeline that progressively transforms simple questions into deep research tasks. Each task is represented as a directed acyclic graph (DAG) of atomic steps and associated checkpoints, enabling the query, DAG, and rubrics to evolve together in a controlled manner. Experiments demonstrate that the benchmark clearly discriminates among models and query types, while its fact-grounded pointwise rubrics enable fine-grained, human-aligned, and stable evaluation. Our data, implementation, and results are publicly available.

The next POI (point-of-interest) recommendation problem is very challenging. It requires not only considering the previous state, location, and user context information, but also analyzing the user behavior. The personalization and uncertainty of user behavior are key issues affecting the accuracy of POI recommendation. Despite this, practice has shown that user behavior is driven by purpose and has certain inherent patterns, which we call behavior inertia. It has a subconscious guiding effect on user decision making. Of course, The effect of this influence will be affected by various factors in real-world scenarios, which we collectively refer to as inertia resistance. Therefore, the POI recommendation model proposed in this paper will fully consider the interaction between behavior inertia and inertia resistance to improve the recommendation accuracy. First, we classify POI according to the purpose of user behavior and construct a purpose prediction model. Second, since we consider the influence of geographic location information on user selection behavior, we construct a prediction model based on POI spatial attributes to calculate the probability of POI check-in under a certain purpose. Meanwhile, to further analyze the influence of behavior inertia on user behavior, we established a prediction model based on user behavior inertia. This model integrates multiple resistance factors such as time, number of POI check-ins, and number of POI categories to explore the influence of behavior inertia on users comprehensively. Finally, the above two models are combined to calculate the probability of a user visiting the POI at the next moment and generate a recommendation list. In order to verify the effectiveness of our method, we conducted extensive experiments on two real-world datasets. The experimental results show that compared to the baseline method, our method improves recall and MAP performance by up to 15% and 20%, respectively.

The performance bottleneck of agents is increasingly shifting from model capability to the robustness of their execution processes. Tools play a central role as the primary interface through which agents interact with external environments, yet existing methods rarely focus on ensuring robust tool use across diverse runtime conditions. To address this problem, we propose ExpG, a mechanism that builds and refines adaptive guidance capturing each tool's capability boundaries and best practices, thereby enabling agents to use tools more robustly and effectively. ExpG consists of three phases: (1) experience acquisition, which analyzes tool invocation quality from historical execution trajectories, producing structured learnable experiences through multi-aspect attribution; (2) experience distillation, which keeps the experience pool effective by filtering unhelpful experiences, selecting representative ones with an equivalence-class-based method, and summarizing them into generalizable guidance; and (3) experience reuse, which applies the guidance adaptively during future task solving. Extensive experiments show that ExpG brings consistent improvements across the tool selection, tool calling, and response generation tasks, enabling smaller agents to outperform larger ones that do not use ExpG. Moreover, ExpG achieves particularly strong gains in challenging settings, suggesting a promising path toward more robust tool use. Our code, experiments, and results are available.

The rapid proliferation of Internet of things devices and the emergence of Multi-access Edge Computing (MEC) have rendered service orchestration and scheduling across diverse devices increasingly complex. Despite OpenHarmony's robust foundation for cross-device collaboration via its high-performance Distributed SoftBus (DSoftBus), current research is deficient in a systematic service-level scheduling mechanism that can concurrently address varied quality-of-service (QoS) requirements and hardware heterogeneity. Moreover, conventional scheduling techniques frequently neglect the affinity traits between services and resources, like preferences for NPU acceleration or lowlatency communication links, while also imposing significant computing overhead. Therefore, we propose AMOS, an affinityaware scheduling method tailored for OpenHarmony services in MEC scenario to tackle these difficulties. AMOS employs a service-centric modeling methodology that quantifies the relationships between services and diverse computing resources, such as CPUs and NPUs, with communication attributes like DSoftBus latency and bandwidth, into a cohesive optimization model. Particularly, it employs an evolutionary search mechanism propelled by a compact elite-micro-population that preserves highquality candidate solutions within a limited group, integrating directed updating with collaborative search strategies to enhance exploration capabilities while efficiently managing computational complexity. Subsequently, we implement a prototype system on a heterogeneous testbed composed of multiple OpenHarmony development boards, such as KHD3863 and KHD3568. Numerical results indicate that, in comparison to notable peer methods, AMOS significantly lowers scheduling overhead and enhances resource utilization, consequently affirming its efficacy and usefulness in affinity-aware service scheduling.

Aim: Existing AI-assisted traditional Chinese medicine diagnostic tools suffer from opaque reasoning processes, passive interaction, and limited treatment plan presentation. This study proposes a knowledge-enhanced visual diagnostic system to improve the transparency and interpretability of syndrome differentiation and treatment. Methods: The system is built upon a Neo4j knowledge graph comprising 241 syndromes, 1,263 symptoms, and 2,485 relations. It incorporates a four-stage symptom matching pipeline (exact, semantic, fuzzy, and large language model verification), an information gain-driven proactive questioning strategy optimized with genetic algorithms, and a multimodal treatment presentation integrating artificial intelligence-generated illustrations, three-dimensional meridian-acupoint models, and evidence-based literature. Results: Knowledge graph constraints reduced non-standard outputs by 32 clinician-assisted diagnosis, and traditional Chinese medicine education. Automated paired-comparison evaluation across 30 cases further demonstrated significant improvements in diagnostic trust (Cohen's d = 1.82, p < 0.001), reduced cognitive load (improvements in four of five dimensions), and higher credibility of evidence-based references (4.21 vs. 2.95). Conclusions: The proposed system enhances the transparency of traditional Chinese medicine diagnostic reasoning and the interpretability of treatment plans through knowledge graph-driven visualization and multimodal interaction, offering a practical solution for trustworthy artificial intelligence-assisted traditional Chinese medicine applications.

Large language models (LLMs) offer new opportunities for recommendation by interpreting item descriptions, user instructions, and external knowledge through natural-language prompts. However, existing graph-augmented LLM recommenders often use knowledge graphs mainly as prompt-level evidence, leaving ranking decisions weakly constrained by structured user-item relations. This is problematic for next-item recommendation, where the model must compare candidates under the same user context while preserving temporal preference, collaborative signals, and attribute matches. To address this issue, we propose GARDRec, a Graph-grounded Adaptive Reasoning and Decision-aware Recommendation framework for LLM-based next-item ranking. GARDRec constructs semantic-structural item representations from textual node features and graph propagation, derives personalized graph contexts from temporally weighted histories and first-order neighborhoods, and aligns graph-derived representations with a frozen LLM through continuous multimodal prompts. Explicit interaction and matching features are injected through late-stage decision branches, while inter-candidate attention and restricted generative likelihood support final ranking. Experiments on three public benchmarks with multiple LLM backbones show that GARDRec generally improves candidate-ranking performance over representative baselines. Ablation and diagnostic analyses verify the contributions of graph projection, neighborhood retrieval, explicit decision features, ranking loss, and generative calibration.

Vision-language models (VLMs) with varying performance and resource requirements are widely deployed, making it difficult for users to select the most appropriate one among numerous VLM candidates. Existing work reveals the performance paradox phenomenon in language models and focuses on routing methods to solve it. However, developing a router for VLM selection is still a critical yet challenging problem, which primarily faces: 1) lack of specialized data, 2) ineffective feature representation, and 3) rigid model space and costly adaptation. In this paper, we construct a multimodal dataset for VLM selection, containing the outputs of seven mainstream VLMs on 32,626 unique image-text queries. We then propose ARMS, a router for VLM selection. ARMS enhances input signals with VLM profiles, employs a simple but effective architecture to improve representations of queries and VLM capabilities. To improve ARMS' adaptation to new VLMs, we propose two extension training strategies: incremental training and independent training. Experimental results on both in-distribution and out-of-distribution test sets demonstrate the effectiveness of ARMS. In particular, using our training strategy, ARMs (only 800M in size) can adapt to a broader VLM space and defeat commercial models like GPT-4o that are hundreds of times larger in scale. Our code, models, and datasets are available in the anonymous repository.

Microservice bad smells (MBSs) affect system quality. However, the lack of comprehensive and detailed explanations of MBSs and current classification methods does not comprehensively encompass the microservice characteristics, making it challenging to study and address MBSs. Existing studies have focused on MBS detection methods, but identifying and quantifying the effects of different smells and properly assessing system health remain challenging. This review aims to provide an exhaustive list of the MBSs and formulate a reasonable classification based on microservice characteristics. By reasonably assessing the harmfulness of smells, we would comprehensively explain different MBSs. We would explore microservice system evaluation methods and propose a microservice system health assessment model (MSHAM) based on MBSs. We conducted a systematic literature review (SLR) of the catalog and evaluation literature in the field of MBSs. We also combined the open questionnaire form to answer the research questions qualitatively and quantitatively. This paper presents a comprehensive list of 69 types of MBSs, incorporating a highly scalable classification. Following the quality characteristics in ISO/IEC 25010:2023, we determined the sets of characteristics affected by different smells, and provided detailed explanations of smells in the form of MBSs knowledge base. Furthermore, we introduced the MSHAM, revealing the quantification process of harmfulness. The formulated smells list and classification method capture MBS characteristics comprehensively, ensuring scalability. We have exposed the impact of MBSs, enhancing researcher and practitioner understanding. MSHAM provides a logical model to quantify MBSs harm and system health, supporting the development of real-time health monitoring.

While knowledge is managed through Wikipedia and software through GitHub, agent skills still lack an infrastructure for large-scale production, governance, and evolution. SkillWiki is a living knowledge infrastructure that supports the organization, grounding, and continuous evolution of agent skills by transforming heterogeneous knowledge into reusable skill assets linked to their originating evidence. Our demonstration presents the complete skill lifecycle, from knowledge ingestion and skill production to provenance-aware exploration, governance, and execution-driven evolution. SkillWiki highlights a future in which knowledge, skills, and execution experience co-evolve within a shared infrastructure. The live demonstration and source code are publicly available at https://github.com/Huangdingcheng/SkillWiki.

Generative recommendation encodes items as hierarchical semantic identifiers (SIDs) and retrieves the next item through autoregressive decoding. Standard next-token prediction, however, does not explicitly cover the multimodal transitions present in interaction sequences, leaving the ground-truth SID vulnerable to irreversible pruning at early beam-search branches. Across three public benchmarks, we find that 91.9%–96.6% of retrieval failures occur within the first two decoding steps. We therefore propose Temporal Autoregressive Alignment (TAAL). During training, TAAL constructs a joint (c_1,c_2) soft target from historical transitions and aligns the early-prefix distribution with a forward KL objective. During inference, it calibrates candidate scores with pointwise mutual information (PMI) to reduce the influence of globally frequent prefixes. On Amazon Beauty, Instruments, and Yelp, TAAL improves NDCG@10 over the standard baseline by 39.5%, 6.7%, and 28.6%, respectively, while increasing full-SID survival by 3.9%–16.6%. Beam-width analysis further shows that the relative survival gain grows as the beam narrows, reaching 39.4% at B=5.

As the volume of scientific submissions continues to grow rapidly, traditional peer review systems are facing unprecedented scalability pressures, highlighting the urgent need for automated reviewing methods that are both scalable and reliable. Existing supervised fine-tuning approaches based on real review data are fundamentally constrained by single-source of data as well as the inherent subjectivity and inconsistency of human reviews, limiting their ability to support high-quality automated reviewers. To address these issues, we propose EchoReview, a citation-context-driven data synthesis framework that systematically mines implicit collective evaluative signals from academic citations and transforms scientific community's long-term judgments into structured review-style data. Based on this pipeline, we construct EchoReview-16K, the first large-scale, cross-conference, and cross-year citation-driven review dataset, and train an automated reviewer, EchoReviewer-7B. Experimental results demonstrate that EchoReviewer-7B can achieve significant and stable improvements on core review dimensions such as evidence support and review comprehensiveness, validating citation context as a robust and effective data paradigm for reliable automated peer review.

Knowledge graph completion (KGC) aims to infer missing facts in knowledge graphs (KGs), thereby improving their completeness and supporting downstream intelligent applications. However, emerging entities and relations in real-world deployments make inductive KGC difficult, especially under few-shot and zero-shot settings. Multimodal information and Large Language Model (LLM)-derived priors can enrich sparse relational contexts, but they may also introduce noisy or hallucinated evidence. To address these issues, we propose DuPLeR, a Dual-Path LLM Reasoning framework for multimodal few-shot KGC. DuPLeR builds a calibrated relation graph by combining multimodal LLM-derived type priors with factual support structures, and performs dual-level structural reasoning over the refined relation topology. Moreover, a dual-pathway multimodal enhancement module regulates message passing with query-relevant multimodal signals and supplements entity representations after graph propagation. Experiments on eight inductive variants of two multimodal KG (MMKG) benchmarks show that DuPLeR achieves robust performance in data-scarce KGC scenarios.

As a structured representation of real-world facts, knowledge graphs (KGs) play a vital role in IoT applications, due to their strong reasoning capabilities and interpretability. However, private user IoT KG data often needs to be centrally collected for embedding training, which poses significant privacy risks and limits the scalability of knowledge-driven downstream applications in distributed IoT environments. Federated learning (FL) has emerged as a promising solution for decentralized model training, eliminating the need for direct data collection. However, existing federated knowledge graph embedding (KGE) methods often struggle to preserve the inherent graph structure of entities and relations, leading to fragmented and incomplete representations. Additionally, they struggle to effectively capture diverse relational dependencies within personal KGs. To address these challenges, this article proposes an enhanced federated KG embedding method for personal knowledge sharing (FPKS) to enable privacy-preserving KGE training. The FPKS framework consists of a central server and multiple federated clients. To enhance entity and relationship alignment across clients, FPKS maintains separate embedding tables for entities and relationships on the server. Moreover, to capture the structural and contextual information of personal KGs, we introduce a local encoder-decoder architecture that employs a graph convolutional network (GCN) variant as an encoder and a KGE scoring function as a decoder. Furthermore, we propose a bidirectional composite operator for GCN (BiDGCN) to enhance multi-relational information aggregation. Extensive experiments on two widely used KG datasets demonstrate that FPKS significantly outperforms existing methods, improving the quality of learned embeddings while ensuring data privacy. Our approach facilitates decentralized personal knowledge sharing, marking an advancement in secure and efficient IoT knowledge-driven services.

Microservices have become a mainstream architectural paradigm, yet microservice bad smells can significantly harm maintainability and performance. Existing detection tools often produce obscure outputs and lack effective integration with runtime observability, making it difficult for operators to interpret results and take timely action. To address this gap, we propose SmellDoc, a customized framework based on Elastic Stack. SmellDoc extends the native observability dashboard with a microservice bad smell detection plugin, integrating detection, knowledge, and health monitoring. It introduces a Custom-Business-Collector to capture business-level metrics, a Re-integration Collector to aggregate heterogeneous runtime data, and detection components that combine static and runtime analyses. SmellDoc supports a knowledge base of 84 smell types and enables detection of 24 representative smells across architectural, runtime, and performance categories. Results are visualized in Kibana through multiple views, providing operators with actionable insights. Case studies on a benchmark microservice system demonstrate that SmellDoc is effective and usable in detecting, visualizing, and analyzing smells, thus enhancing runtime observability and accelerating troubleshooting to maintain a high level of Quality of Service.

Industrial digital twin (IDT) technology builds high-fidelity virtual representations of physical systems to support real-time synchronization, predictive analytics, and intelligent decision-making across the lifecycle of industrial assets. As a core enabler of Industry 5.0 and intelligent manufacturing initiatives, IDT enhances operational efficiency, reduces resource consumption, and strengthens system resilience by integrating sensing, modeling, simulation, and control within a unified cyber–physical framework. This review summarizes key advances in IDT research. It outlines enabling technologies with a focus on multi-scale modeling, real-time data fusion, high-performance simulation, artificial intelligence (AI) and machine learning analytics, and extended-reality (XR)-assisted visualization. Representative applications are discussed in areas such as intelligent manufacturing, equipment health management, supply-chain optimization, and energy-efficient operations, supported by industrial case evidence. Current challenges include fidelity assurance, heterogeneous data management, system integration complexity, organizational adaptation, and quantitative evaluation of investment value. Existing solution strategies are synthesized from both technological and management perspectives. Emerging trends are highlighted, including the integration of IDT with AI foundation models, quantum and edge computing, next-generation communication networks, and blockchain-based trust infrastructures. These developments are expected to broaden application domains and shape future industrial ecosystems.

In the metaverse, human-agent interactions increasingly take diverse forms, such as natural dialogues, which enhance user experience but pose challenges for precise service delivery, as users often provide incomplete or ambiguous information-a phenomenon we term fuzzy requirements. This paper addresses how to clarify fuzzy requirements through multi-turn interactions. We formally define the task, construct a benchmark across five domains with user profiles, and propose Tower, a user experience-oriented decision-making algorithm for fuzzy requirement clarification. Tower organizes service parameters into a feature tree as a search space for inquiries and refines it using user profiles and diverse interview questioning strategies. To enable realistic evaluation, we develop a user simulator based on action selection, modeling user personality traits using prospect theory and Fitts's law from cognitive psychology. The simulator proactively presents fuzzy requirements and responds adaptively according to personality traits. Using dialogue histories between Tower and the simulator, we train TowerLM, an automatic clarification model with strong domain generalizatio. Experimental results show that both Tower and TowerLM achieve state-of-the-art performance across three user personality types and can be seamlessly integrated as plugins into existing service delivery methods, improving the delivery success rate under fuzzy requirements on four public datasets.

LLM-based agents now have strong general capabilities. However, they still struggle with domain-specific tasks, motivating the integration of external tools to broaden their capabilities. The open-source community offers a vast array of AI models typically released as heterogeneous research artifacts, whereas transforming them into ready-to-call APIs is costly and labor-intensive. Automated model deployment is therefore essential for bridging the gap between model resources and tool usability, yet it remains a long-horizon, multi-stage task that has not been sufficiently explored. To tackle this challenge, we introduce Model Automated Deployment Engine (MADE), a dual-agent coordination system. Specifically, given a model resource, MADE iteratively constructs and validates the deployment artifacts, updates its deployment belief based on execution feedback, and revisits invalid upstream artifacts until the model is successfully served as a ready-to-call API that can then be used by other agents. We further introduce M2ABench, a benchmark for the task of transforming Models to ready-to-call APIs. M2ABench comprises 122 real-world models with standardized test cases for evaluation. Experimental results demonstrate that MADE achieves a deployment success rate of 68.85

Large Language Models (LLMs) exhibit impressive capabilities yet suffer from sensitivity to slight input context variations, hampering reliability. Conventional metrics like accuracy and perplexity fail to assess local prediction robustness, as normalized output probabilities can obscure the underlying resilience of an LLM's internal state to perturbations. We introduce the Token Constraint Bound , a novel metric that quantifies the maximum internal state perturbation an LLM can withstand before its dominant next-token prediction significantly changes. Intrinsically linked to output embedding space geometry, provides insights into the stability of the model's internal predictive commitment. Our experiments show correlates with effective prompt engineering and uncovers critical prediction instabilities missed by perplexity during in-context learning and text generation. offers a principled, complementary approach to analyze and potentially improve the contextual stability of LLM predictions.

The construction of domain knowledge graphs is a key strategy for achieving industrial intelligence and promoting collaborative data intelligence within industries. Relational databases in enterprise information systems contain vast amounts of business data, which serve as a crucial source of industry knowledge. However, traditional rule-based ontology extraction methods often perform poorly when applied to enterprise databases with diverse design paradigms and expanding scales. Moreover, ontologies generated from different databases may vary in structure, which poses new challenges to achieving cross-enterprise data collaboration. As a result, designing an efficient and high-quality automated ontology extraction method remains a major challenge. To this end, we propose a domain knowledge extraction approach for relational databases (KE4RDB). First, we design a hierarchical three-layer ontology model that not only reduces the cost of mapping relational data to ontologies but also improves the efficiency of knowledge extraction. Based on this model, we develop a method for extracting knowledge from relational databases and integrating it into unified domain ontologies. Finally, we conducted experiments on three domain-specific databases, demonstrating that the KE4RDB performs excellently in domain knowledge extraction and ontology generation tasks.

In the metaverse, human-agent interactions increasingly take diverse forms, such as natural dialogues, which enhance user experience but pose challenges for precise service delivery, as users often provide incomplete or ambiguous information—a phenomenon we term fuzzy requirements. This paper addresses how to clarify fuzzy requirements through multi-turn interactions. We formally define the task, construct a benchmark across five domains with user profiles, and propose Tower, a user experience-oriented decision-making algorithm for fuzzy requirement clarification. Tower organizes service parameters into a feature tree as a search space for inquiries and refines it using user profiles and diverse interview questioning strategies. To enable realistic evaluation, we develop a user simulator based on action selection, modeling user personality traits using prospect theory and Fitts's law from cognitive psychology. The simulator proactively presents fuzzy requirements and responds adaptively according to personality traits. Using dialogue histories between Tower and the simulator, we train TowerLM, an automatic clarification model with strong domain generalizatio. Experimental results show that both Tower and TowerLM achieve state-of-the-art performance across three user personality types and can be seamlessly integrated as plugins into existing service delivery methods, improving the delivery success rate under fuzzy requirements on four public datasets.