While user preferences are important to cross-domain recommendation (CDR), existing methods primarily discover preferences under specific, yet possibly redundant, item features. To this end, we first propose a novel Preference Prototype-Aware (PPA) learning method to quantitatively learn user preferences while minimizing disturbances from the source domain. It introduces a mix-encoder and a proto-decoder. On the one hand, the mix-encoder learns better general representations of interacted items and captures the intrinsic relationships between items across different domains. On the other hand, the proto-decoder implements a learnable prototype matching mechanism to quantitatively perceive user preferences, avoiding disturbances caused by item features from the source domain. Moreover, through experiments on PPA, we observe another two issues that affect existing CDR methods’ performance, i.e., the semantic deficiency caused by sparse item categories and the imbalance weights caused by different user-item distributions. Thus, we further propose a LoRA-based extractor and a domain cross-attention module to alleviate the two issues, respectively. The PPA incorporating with new extractor and attention module is called PPA++. Extensive experiments show that PPA++ outperforms the other state-of-the-art counterparts in four different CDR scenarios.

Large language models (LLMs) have demonstrated significant potential in complex question answering (QA) by reasoning on knowledge graphs (KGs), which offer structured external knowledge. Existing methods for KG-augmented reasoning, including iterative and once paradigms, have enhanced LLMs’ performance in complex QA and emphasized the significance of reasoning through KG relations. However, a significant challenge remains in constructing optimal reasoning plans from KG relations while maximizing their reasoning capabilities. To address this challenge, we propose Graph-of-Thought Guided Reasoning (GoT-R), which enhances LLMs for complex QA. GoT-R integrates three components: (i) atomic relations selection, which uses an encoder to identify relevant relations; (ii) graph-of-thought (GoT) construction, which builds a graph-of-thought combining KG-aligned relations and LLM’s inherent knowledge; and (iii) GoT-guided hybrid retrieval and reasoning, which integrates relation-aware and relation-unaware retrieval to generate reliable and comprehensive reasoning evidence. By leveraging KG relations and LLM knowledge, GoT-R achieves faithful and interpretable reasoning while reducing reliance on high-quality QA datasets. Experimental results on four complex QA datasets demonstrate GoT-R’s effectiveness and generalizability. The source code for this project is available at https://github.com/Peixuan-Huang/GoT-R .

The growing emphasis on data-driven decision-making has created an increasing demand for effective dataset search. In the spatial data domain, the need becomes even more prominent, as analyzing patterns and predicting the behaviors of entities such as individuals or vehicles often require discovering relevant datasets. However, most existing spatial dataset search methods are primarily designed for scenarios that focus on geographic proximity, overlooking cases where datasets from different regions share similar distribution patterns. In this paper, we propose a Distribution Pattern similarity-based Spatial dataset Search scheme (DPSS), which leverages a probability distribution-based similarity model to quantify the similarity between datasets. By incorporating kernel density estimation, our model significantly enhances search effectiveness in identifying datasets with similar spatial distributions. To further improve efficiency, we design an accelerated similarity computation method and a dataset indexing mechanism to filter candidates, leading to an optimized version, DPSS+. Extensive experiments conducted on three real-world repositories containing 100,000 spatial datasets demonstrate that DPSS+ achieves high search accuracy while requiring only around 10 s to complete a query.

In the data-driven era, dataset search has become a critical task in data science and engineering. Traditional spatial dataset search methods primarily rely on keyword or range queries, which are insufficient for capturing user intent expressed through exemplar datasets. To address this gap, this paper investigates the problem of top-k spatial dataset search using exemplar datasets as input. A novel grayscale image-based similarity model is first proposed, which maps the spatial distribution of datasets into grayscale images to capture detailed distribution features. Based on this model, a baseline search scheme (GIDS) is proposed. To further improve the search efficiency, an optimized search scheme (GIDS +) is introduced, which incorporates two key optimization strategies: a Morton code-based strategy to accelerate similarity calculations and a ω -MSDtree-based strategy to enable efficient pruning during candidate filtering. Experiments conducted in two real-world spatial data repositories demonstrate that the proposed methods outperform existing approaches in search efficiency, providing a new solution for spatial dataset search.

To ensure real-time performance and reliability, workloads in an airborne-cloud are often planned in advance, thereby making key task attributes—such as estimated arrival time, execution duration, and resource requirements—available prior to scheduling. Such foresight provides an opportunity to optimize scheduling from a long-term perspective, whereas most existing airborne-cloud schedulers are designed for online settings and assume that future task information is unavailable. To bridge this gap, we formalize the planned task scheduling problem and propose ElitePT (Elite Genetic Algorithm for Planned Tasks), a scheduling strategy that explicitly exploits planned information. ElitePT evaluates candidate schedules with a long-term performance objective that jointly accounts for load balancing and energy consumption over a planning horizon, and it strengthens the evolutionary search by injecting heuristic-constructed elite individuals while enforcing population diversity through Hamming-distance-based initialization. Experiments in CloudSim show that, compared with representative baselines including least-loaded scheduling, rotating scheduling, and a standard genetic algorithm, ElitePT delivers consistently better long-term scheduling quality and stability; across diverse cluster scales, heterogeneity settings, and overlap levels, it improves long-term load balancing by 20–50

Although large language models(LLMs) have been applied across various fields and achieved remarkable success, concerns have also been raised regarding the potential for generating harmful content. Jailbreak, an emerging research direction, aims to bypass the safety mechanisms of LLMs and induce undesired responses. Research on jailbreak attack methods can reveal potential safety risks in LLMs and better guide researchers in developing corresponding defense strategies. However, many existing attack methods either require access to the internal structure of the target model, or incur high costs due to the need to design complex nested scenarios. We propose a general and efficient jailbreak framework for LLMs that integrates knowledge graph(KG), called HBS-KGLLM, which consists of three main components: (1) Harmful Behavior Substitution, (2) KG template nesting, and (3) KG-to-text conversion. We evaluated HBS-KGLLM on five SOTA LLMs, including both open-source and closed-source models. Extensive experiments demonstrate that HBS-KGLLM significantly improves the attack success rate compared to existing baselines, while also greatly reducing both time and financial costs. Using GPT-4 as an example, our method improves the ASR by nearly 25

Session-based recommendation systems aim to predict user clicks using anonymous session data. Current models struggle with understanding complex item transition patterns and are affected by noise in the data, thus reducing accuracy. To address these problems, we introduce a self-supervised dual graph and intention association technique for session-based recommendations, named SDGIA (Self-supervised Dual Graph and Intention Association). SDGIA constructs a global undirected graph and session-directed graphs, enhancing information representation to capture transition patterns within and across sessions. A self-supervised learning mechanism improves feature extraction and generalization, meanwhile an intention association module filters out noise for more precise item representations. Experiments on three datasets demonstrate that SDGIA significantly outperforms existing models.

Speech editing for content creation requires precise control over both what an edit should do and where it should apply. Free-form natural language provides a flexible interface for expressing edit requests, but its ambiguity may leave the intended operation, parameters, or target region underspecified. We study a precise and explicit interface for speech editing: a transcript-grounded structural edit instruction with XML-style tags explicitly specifies typed operations and localizes them to transcript spans or boundaries. This semantic timeline avoids explicit timestamp alignment and provides an externally inspectable contract for compositional edits. We instantiate the interface in dots.tts.edit, an editor adapted from the continuous autoregressive dots.tts foundation model. Four representative speech-creation controls cover lexical content, affective expression, pitch and speaking-rate delivery, and temporal phrasing through text, emotion, prosody, and pause editing. Task-specific data pipelines construct operation- and scope-controlled pairs while retaining source-derived context outside each target region. We further introduce doteBench, a bilingual evaluation suite that measures precise instruction following, local preservation, and audio quality across the four controls and their composition. Experiments show leading overall instruction following and local preservation across its five editing categories, while audio quality remains comparable to existing open-source systems. Across three Seed-TTS-Eval shards, the model shows negligible differences from the base model in zero-shot TTS recognition error rate and speaker similarity.

Cross-domain recommendation (CDR) improves target-domain recommendation by leveraging behavioral signals from auxiliary domains, alleviating data sparsity and cold-start issues. Universal cross-domain recommendation (UniCDR) further learns a single model to unify multi-domain preferences and support both dual-domain and multi-domain scenarios. However, incorporating source-domain information can even hurt target-domain performance, a phenomenon known as negative transfer. This is because cross-domain signals are not uniformly beneficial: representation learning may capture shortcut cues that break under domain shift, and static fusion may further propagate such unreliable components into the target representation. To explicitly address what to transfer, we propose the Refined Transfer Framework (RTF), an end-to-end architecture that decouples robust representation learning from selective cross-domain fusion. At the representation layer, RTF employs a temperature-scaled attention encoder over the source history and the global history, together with popularity down-weighting to suppress shortcut cues. At the fusion layer, we introduce a domain-routed decoder that aligns multi-source representations in a shared space and selectively fuses auxiliary signals with a global fallback. Finally, we optimize RTF with recommendation supervision and dual contrastive objectives to enhance intra-view robustness and cross-view consistency. Extensive experiments on Amazon benchmarks across four UniCDR scenarios demonstrate that RTF yields competitive and stable improvements over strong baselines.

Recent advancements in Large Language Models (LLMs) have significantly extended context windows, igniting Haystack (NIAH) framework, systematically evaluates LLMs and RAG methods in controlled long-context settings. It extends beyond traditional NIAH by incorporating more practical and complex scenarios like multineedle, long-needle, and needle-in-needle configurations and leveraging the synthetic dataset to mitigate LLM biases. The experiments aim to address three research questions in long-context scenarios: (1) performance tradeoffs between LLMs and RAG, (2) error patterns in RAG, and (3) RAG's limitations in complex settings. Results show that smaller LLMs benefit more from RAG. In all settings, RAG achieves a win rate of 82.58% over direct answers. Additionally, it is found that retrieval noise and chunk ordering degrade RAG performance, and we further summarized typical error patterns, including omissions due to noise, hallucinations under high noise critical conditions, and self-doubt behaviors, as well as how these phenomena vary with context length. Finally, in some challenging scenarios, experiments show that deep reasoning models are more easily affected by distractors. These findings highlight the complementary roles of RAG and LLMs and offer actionable insights for optimizing deployment strategies (https://github.com/Tongji-KGLLM/U-NIAH).

Next Point-of-Interest(POI) recommendation focuses on predicting the next location a user would visit, where the demands of users depend on the implicit user preference implied by the past behavioral trajectories and are influenced by future behavior factors as well. However, most existing methods that consider user future preference merely focus on the fine-grained POI check-in trajectories of single users' sequential migrations. This overlooks the potential benefits of incorporating coarse-grained signals (global check-ins) to enhance user preference learning. And yet, traditional global coarse-grained structures are mainly constructed by graph neural networks and fixed graph structures to obtain POI representations that aggregate information and capture node correlations, which can move away from the optimal graph structure due to noise and adaptation issues, leading to sub-optimal performance of recommendations. In this paper, we propose a framework based on Adaptive Graphs Learning and Future Preferences(AGFP), which develop an adaptive POI graph matrix and learns it via similarity learning with POI embeddings, after which we merge the learning POI representation with the side information and spatio-temporal information, then generate the user's future preferences through the fusion of long-term preference encoder with the short-term preference encoder in a self-ensembling manner. Our model is extensively experimented on three real-world datasets. The experimental results demonstrate that AGFP shows superior performance.

Retrieval Augmented Generation (RAG) has emerged as a promising solution to address hallucination issues in Large Language Models (LLMs). However, the integration of multiple retrieval sources, while potentially more informative, introduces new challenges that can paradoxically exacerbate hallucination problems. These challenges manifest primarily in two aspects: the sparse distribution of multi-source data that hinders the capture of logical relationships and the inherent inconsistencies among different sources that lead to information conflicts. To address these challenges, we propose MultiRAG, a novel framework designed to mitigate hallucination in multi-source retrieval-augmented generation through knowledge-guided approaches. Our framework introduces two key innovations: (1) a knowledge construction module that employs multi-source line graphs to efficiently aggregate logical relationships across different knowledge sources, effectively addressing the sparse data distribution issue; and (2) a sophisticated retrieval module that implements a multi-level confidence calculation mechanism, performing both graph-level and node-level assessments to identify and eliminate unreliable information nodes, thereby reducing hallucinations caused by inter-source inconsistencies. Extensive experiments on four multi-domain query datasets and two multi-hop QA datasets demonstrate that MultiRAG significantly enhances the reliability and efficiency of knowledge retrieval in complex multi-source scenarios. \textcolor{blue}{Our code is available in https://github.com/wuwenlong123/MultiRAG.

Recent advancements in Large Language Models (LLMs) have expanded their context windows to unprecedented lengths, sparking debates about the necessity of Retrieval-Augmented Generation (RAG). To address the fragmented evaluation paradigms and limited cases in existing Needle-in-a-Haystack (NIAH), this paper introduces U-NIAH, a unified framework that systematically compares LLMs and RAG methods in controlled long context settings. Our framework extends beyond traditional NIAH by incorporating multi-needle, long-needle, and needle-in-needle configurations, along with different retrieval settings, while leveraging the synthetic Starlight Academy dataset-a fictional magical universe-to eliminate biases from pre-trained knowledge. Through extensive experiments, we investigate three research questions: (1) performance trade-offs between LLMs and RAG, (2) error patterns in RAG, and (3) RAG's limitations in complex settings. Our findings show that RAG significantly enhances smaller LLMs by mitigating the "lost-in-the-middle" effect and improving robustness, achieving an 82.58 observe that retrieval noise and reverse chunk ordering degrade performance, while surprisingly, advanced reasoning LLMs exhibit reduced RAG compatibility due to sensitivity to semantic distractors. We identify typical error patterns including omission due to noise, hallucination under high noise critical condition, and self-doubt behaviors. Our work not only highlights the complementary roles of RAG and LLMs, but also provides actionable insights for optimizing deployments. Code: https://github.com/Tongji-KGLLM/U-NIAH.

Graph Neural Networks (GNNs) have emerged as the preferred backbone model of recommender systems, credited to their strong capability in capturing the intricate topological relationships within user-item interactions. Nevertheless, a common oversight in existing studies is the presumption of the inherent reliability of these interactions, ignoring the reality that a significant fraction of useritem engagements, such as accidental clicks, are inherently noisy. Extensive studies have revealed that, GNN is vulnerable to such noisy edges within the graph-structured data, as those noisy edges can mislead the network into overfitting incorrect patterns of interactions, thereby propagating such incorrect information through entire interaction network. To address those challenges, in this paper, we propose a novel noise-robust GNNs based training strategy for recommendation, known as Noise-Resistant Graph Collaborative Filtering (NR-GCF). NR-GCF innovatively adopts a two-stage learning paradigm to filter out unreliable interactions, leveraging the memorization effect of GNNs. It further utilizes representation modulation to learn noise-resistant embeddings, enhancing robustness for recommendation tasks. Comprehensive experiments and ablation studies demonstrate the effectiveness and robustness of the proposed NR-GCF. Our implementation has been made available in the attachment: https://github.com/1197151063/NRGCF.git

Knowledge Graph (KG) faces challenges of incompleteness, driving the requirements of Knowledge Graph Completion (KGC). The development of Large Language Models (LLMs) provides a new perspective for KGC research. Several methods instruct LLMs to conduct KGC by Prompt Engineering. However, they struggle with missing entity/relation descriptions, “text mismatch” between LLMs' responses and entities in KG, and insufficient utilization of structural information. To address the above challenges, we propose RGM-KGC, a multi-stage method for LLMs-based KGC. It first retrieves a subgraph related to query, with the aim to provide structured information of entities for LLMs. Subsequently, in order to enrich the semantics of entities and relations, RGM-KGC guides LLMs via CoT prompts in generating additional knowledge based on the query and its subgraph, thereby enhancing LLMs' cognition. Finally, RGM-KGC conducts similarity matching between LLMs' responses and the entities in knowledge graph to obtain candidate entities, with the aim to solve “text mismatch” problem. We conduct comprehensive experiments on the link prediction task using two benchmark datasets, FB15K-237-N and Wiki27K, which proved the effectiveness of RGM-KGC.

Intelligent question-answering (QA) systems powered by large language models (LLMs) have shown significant promise across various fields, but several challenges still hinder their application in mechanical manufacturing. This domain involves complex textual knowledge and numerous process design parameters, often presented in tabular form, which traditional LLMs struggle to extract effectively. We propose a novel intelligent QA system, MAR, tailored to address this in the mechanical manufacturing domain. MAR integrates open-source LLMs, Retrieval-Augmented Generation (RAG), and knowledge Re-ranking techniques. By combining RAG with Re-ranking, the system significantly enhances the accuracy and efficiency of knowledge retrieval, overcoming the limitations of traditional RAG methods. The system utilizes the LoRA+-fine-tuned Qwen2-7B model alongside a knowledge base containing industry standards and process design data. Experimental results show that the combination of RAG and Re-ranking outperforms traditional LLMs, as evidenced by improvements in BLEU-4 and ROUGE metrics. This research provides a practical solution for intelligent QA in mechanical manufacturing, effectively handling complex knowledge and precision-based tasks.

Heterogeneous hypergraph neural networks are powerful tools to capture complex correlations among various nodes in Heterogeneous Information Networks (HINs). Despite satisfied performances of them, they are still plagued by the following problems: (1) They cannot capture the correlations in structural and semantic view at once, leading to topological information loss. (2) Due to the number of nodes being greater than the number of node types, node-level self-attention they used causes massive parameters and leads to high time consumption. (3) Interactions in meta-paths may be redundant, resulting in the correlations bias. To address the three issues, we propose an efficient M ulti- V iew H eterogeneous H yper g raph C onvolutional N etwork (MVH $^{2}$ GCN). It first constructs relational and semantic hypergraphs based on different types of edges and meta-paths respectively, to represent the complex correlations in structural view and semantic view. Meanwhile, the clean semantic hypergraphs are generated by structure learning network to avoid redundancy. Then, an efficient hypergraph convolutional network is designed to learn node embeddings. By doing so, correlations in the two views are captured. Finally, the learned node embeddings from two views are aggregated via a gated embedding fusion module for downstream tasks. Experiment results demonstrate that MVH $^{2}$ GCN is effective and efficient.

A well-designed recommender system can accurately learn the embeddings of users and items, reflecting the unique preferences of users. Traditional recommendation techniques usually focus on modeling the singular type of behaviors between users and items. However, in many practical recommendation scenarios (e.g., social media, e-commerce), there exist multi-typed interactive behaviors in user-item relationships, such as click, tag-as-favorite, and purchase in online shopping platforms. Thus, how to make full use of multi-behavior information for recommendation is of great importance to the existing system, which presents challenges in two aspects that need to be explored: (1) Utilizing users’ personalized preferences to capture multi-behavioral dependencies; (2) Dealing with the insufficient recommendation caused by sparse supervision signal for target behavior. In this work, we propose the Knowledge Enhancement Multi-Behavior Contrastive Learning framework (KMCLR)[54], including two Contrastive Learning tasks and three functional modules to tackle the above challenges, respectively. In particular, we design the multi-behavior learning module to extract users’ personalized behavior information for user-embedding enhancement, and utilize knowledge graph in the knowledge enhancement module to derive more robust knowledge-aware representations for items. In addition, in the optimization stage, we also model the coarse-grained commonalities and the fine-grained differences between multi-behavior of users to further improve the recommendation effect, and propose a joint training paradigm to enhance the learning effect of KMCLR in the joint learning module. Besides, we also considered how to make full use of temporal signals to enhance the effectiveness of multi-behavior recommendations in scenarios with time information and designed a novel encoder to address this issue. Extensive experiments and ablation tests on the three real-world datasets indicate our KMCLR outperforms various state-of-the-art recommendation methods and verify the effectiveness of our method.

Session-based recommendation (SBR) focuses on forecasting the next item a user is likely to select using brief and anonymous sequences of interactions. Existing methods face three key challenges: (1) difficulty in distinguishing noisy transitions within sessions, (2) absence of explicit modeling for target intent, and (3) misalignment between intra- and inter-session information. We propose TiDGRec (Target-intention aware Dual-Graph Recommender), a framework designed to address these limitations through hierarchical denoising and target-guided dual-graph learning. A Target Proxy Node (TPN) is introduced into the Sequential Transition Graph (STG) to capture user intent representations. An Adaptive Target-aware Sparsifier (ATS) based on dynamic α _s , adaptively filters irrelevant transitions. The learned target representation and item embeddings from the Cross-session Co-occurrence Graph (CCG) are jointly input to the Target-guided Cross-graph Filter (TCF) to enhance target-aware global relations. By connecting STG and CCG through shared target signals, TiDGRec forms a dual-graph, dual-target architecture that enhances intent alignment, suppresses semantic noise, and improves overall recommendation quality. A comprehensive evaluation across various benchmark datasets demonstrates that TiDGRec achieves superior performance compared to existing SBR methods.

Recommender systems aim to personalize user experiences by suggesting relevant items based on historical interactions and preferences. However, traditional CoT reasoning struggles to align generative agents' outputs with real-time user preferences, especially when faced with sparse or changing data. In this paper, we introduced the MetaCoT-A*-MCTS framework by utilizing A* for deterministic pathfinding and MCTS for stochastic exploration, which combines the strengths of A* search and MCTS to improve personalized recommendation systems. The hybrid approach allows for dynamic exploration of reasoning paths, optimizing the balance between computational efficiency and recommendation quality. Our experiments on datasets including MovieLens1M, MovieLens-10M, Amazon-Book, and Steam demonstrate that that Meta-CoT-A*-MCTS consistently outperforms other methods, achieving higher F1 scores across various benchmarks and model sizes. This framework effectively balances high-quality recommendation generation with computational efficiency, reducing time complexity while maintaining high-quality recommendations.