While Large Language Models (LLMs) offer promise for automating SQL query rewriting, their ”black-box” nature presents a major obstacle to practical adoption. The opacity of their reasoning leads to a lack of trust, unexplainable logic, and a high risk of generating syntactically or semantically flawed queries. Furthermore, because of search space pruning and strict cost models, classic database optimizers, despite their robustness, frequently struggle to handle the intricate structural transformations needed for contemporary analytical workloads. We provide Blueprint-SQL, a novel framework that reframes rewriting from opaque text production to transparent, organized planning, in order to address these issues. Our approach trains an LLM agent via reinforcement learning to produce a ”rewrite blueprint”—a high-level, human-readable sequence of composable actions like CONVERT_SUBQUERY_TO_CTE or REORDER_JOIN. This blueprint is then deterministically executed by applying each action to the query’s Abstract Syntax Tree (AST), a process that ensures the final output remains syntactically valid within the supported action space. This architecture decouples strategic reasoning from low-level code generation, yielding three key advantages: built-in explainability, guaranteed robustness, and learned adaptability. Through extensive evaluation against production-grade RDBMS optimizers (PostgreSQL), hybrid LLM-rule systems, and black-box generative models, we demonstrate that Blueprint-SQL achieves state-of-the-art performance on standard benchmarks. Critically, it maintains 100
The identification of anomalies in time series has gained significant attention in both academia and industry, leading to the development of numerous detection approaches. However, anomaly definitions vary across applications, multiple anomaly types may coexist, and the dynamic nature of monitored data further complicates accurate detection. To address these challenges, we propose an explainable, contextualized anomaly detection strategy that can be applied on top of any anomaly detector. A key element of this strategy is the TECAD framework for context generation, which incorporates heterogeneous data sources and leverages causal discovery to derive the context in which an anomaly detector operates. TECAD is model-agnostic, designed for online settings, and serves both explainability and decision-making enhancement in anomaly detection. Experimental results demonstrate that TECAD can function as an explainability method, improving the conciseness and consistency of explanations compared to state-of-the-art approaches, while also providing interpretations beyond monitored features. Furthermore, applying TECAD for contextualized decision-making can significantly enhance the effectiveness of anomaly detectors, yielding improvements of up to a twofold increase in F1-scores, as evidenced by its evaluation in three challenging real-world use cases that are characterized by data and concept drifts.
Knowledge Graphs (KGs) cover abundant factual information and semantic associations between items. They are often used to enhance the representation ability of models in the field of recommendation systems, and provide key support for improving recommendation performance. However, not all information in KGs has the same relevance and value to the target recommendation task. KGs themselves generally have significant inherent sparsity. Some connection information obtained from KGs may also introduce extra noise. Moreover, user behavior collaborative signals and KGs semantic signals are often modeled and processed separately. In this paper, we propose a new knowledge graph-based recommendation framework called KMDCL. To fully explore the semantic value of KGs and improve their adaptability to recommendation tasks, we build a knowledge graph completion strategy using a locally deployed large language model (LLM) . We also propose a knowledge graph denoising method based on diffusion models to accurately filter out low-confidence noisy relations. These two methods are combined to purify and enhance the information of KGs. On this basis, we combine the optimized KG with the user-item interaction graph to build a multi-view representation space for items. We also introduce a contrastive learning mechanism to enhance the semantic discrimination and expression ability of item embeddings. Our extensive experiments on three public datasets show that KMDCL has achieved significant performance improvement compared with a variety of baseline models. Experimental results show that this method can effectively alleviate the challenges brought by data sparsity and knowledge graph noise, and enhance the learned representation to improve the recommendation performance.
Metaphorical expressions evoke stronger emotions than literal language. However, their implicit sentiment mapping characteristics pose significant challenges for sentiment analysis. Existing studies exhibit two major limitations: first, most models focus on modeling the partial of three core elements (the source-domain word S, the target-domain word T, and the metaphorical context C), neglecting the interactions among them in metaphors; second, current research in computational metaphor predominantly adopts a single-type metaphor pattern, making it difficult for models to capture shared representations across different metaphor types, thereby limiting their generalization. To address these issues, this study proposes a multi-type metaphor sentiment analysis framework that jointly models S, T, and C, aiming to accurately identify the sentiment of complex metaphorical texts while revealing the path of metaphorical sentiment transfer, thereby facilitating sentiment analysis of metaphor-rich texts. To effectively model the three elements, a metaphor sentiment analysis model integrating Multi-Task Contrastive Learning and Instruction Tuning (MTCL-IT) is proposed. Finally, multi-perspective experiments are conducted on the constructed English dataset EMSA and Chinese dataset CMSA, both containing S, T, sentiment polarity, and metaphor type information, fully validating the effectiveness of the proposed three-element joint modeling framework and model in sentiment identification and understanding of sentiment mapping mechanisms.
Structured multi-agent debates among Large Language Models (LLMs) have emerged as a powerful paradigm for enhancing reasoning reliability and argumentative coherence. Motivated by the European Space Agency’s (ESA) interest in trustworthy AI for space operations, this study proposes a moderated, domain-adaptive multi-agent debate framework applied to the high-stakes domain of satellite communications (SatCom). Specifically, it assesses (i) the efficacy of structured deliberation against single-agent baselines, and (ii) the impact of model heterogeneity versus homogeneity. A single-agent baseline is compared against a multi-agent framework deploying a moderator and two domain-specialized experts. These systems utilize local 70B-parameter LLMs in homogeneous (Llama-3.3) and heterogeneous (Llama-3.3 + DeepSeek-R1 + Qwen-2.5) configurations, all augmented with a shared, curated Retrieval-Augmented Generation (RAG) corpus combining academic institutional sources and ESA material from the Nebula portal (SatNex V programme). Outputs from 213 technical queries are evaluated via LLM-as-a-judge across three phases: baseline proficiency, strategic reasoning, and executive readiness. Single-agent systems lead in encyclopedic tasks, where retrieval suffices over deliberation. However, both multi-agent configurations outperform in strategic reasoning, with heterogeneous debates achieving superior performance in executive scenarios by victory margins of up to 2.75 points on a 10-point scale. These results validate architectural diversity as a decisive factor in resolving high-complexity technical conflicts. Ultimately, this work delivers a generalizable, fully traceable deliberation framework suitable for real-world, mission-critical environments. Code, prompts, and evaluation data are publicly available at https://github.com/amozo-es/multi-agent-debate/ .
Modern retrieval pipelines compose lexical, dense, structural, agentic, and reranking strategies, yet routing each query to its best paradigm remains elusive. We document a routing paradox: across six benchmarks and a five-strategy retrieval stack, no learned hard router beats the strongest fixed paradigm. To investigate why, we formalize a predictive ceiling—the irreducible conditional entropy of the oracle paradigm given query features—and show empirically that, even with rich corpus-aware and probe features, the mutual information between features and the oracle paradigm accounts for only a small fraction of total routing entropy, capping deterministic paradigm-selection accuracy well below the oracle. We then introduce CASR (Corpus-Aware Soft Routing), an algorithmic method that hedges under uncertainty by predicting per-paradigm performance signatures via multi-output regression and converting them into temperature-softmaxed fusion weights. CASR matches strong fusion baselines in retrieval quality, significantly outperforms every individual paradigm including the strongest, and uniquely enables cost-aware selective inference, halving paradigm invocations at negligible accuracy cost by skipping paradigms whose predicted weight falls below a threshold. RCT-MARS (Retrieval Complexity Taxonomy through Multi-paradigm Algorithm Routing and Selection) provides the empirical taxonomy, the formal predictive-ceiling diagnosis, and the CASR algorithmic method; we release code, indices, and a practitioner decision tree. All quantitative results, statistical tests, and confidence intervals are reported in the main paper.
Multi-hop question answering (MHQA) requires aggregating information across multiple heterogeneous evidence passages, simulating human step-by-step reasoning and clue tracing to construct a complete logical path toward the answer. However, existing approaches face key limitations: chain-of-thought paradigms are prone to error accumulation across reasoning steps, while tree-structured paradigms often struggle with efficiency due to the vast discrete search space and slow convergence. To address these challenges, we propose the PMA framework, which integrates a prior estimate of the number of reasoning steps with structurally constrained search to achieve efficient and stable reasoning using minimal yet sufficient paths. Specifically, PMA consists of three core components: (1) a fine-tuned BERT-based hop predictor to estimate the required number of reasoning steps, guiding the scope of tree expansion; (2) a structurally constrained Monte Carlo Tree Search (MCTS) that regulates search depth and dynamically selects expansion nodes using coherence and adaptability scores; and (3) a multi-path answer aggregation module that consolidates candidate answers via majority voting, with ties resolved by selecting the path with the highest average score. Experimental results demonstrate that PMA achieves competitive accuracy while significantly reducing token consumption, highlighting its advantages in both efficiency and robustness.
Popular word embedding methods such as GloVe and Word2Vec are related to the factorization of the pointwise mutual information (PMI) matrix. In this paper, we establish a formal connection between correspondence analysis (CA) and PMI-based word embedding methods. CA is a dimensionality reduction method that uses singular value decomposition (SVD), and we show that CA is mathematically close to the weighted factorization of the PMI matrix. We further introduce variants of CA for word-context matrices, namely CA applied after a square-root transformation (ROOT-CA) and after a fourth-root transformation (ROOTROOT-CA). We analyze the performance of these methods and examine how their success or failure is influenced by extreme values in the decomposed matrix. Although our primary focus is on traditional static word embedding methods, we also include a comparison with a transformer-based encoder (BERT) to situate the results relative to contextual embeddings. Empirical evaluations across multiple corpora and word-similarity benchmarks show that ROOT-CA and ROOTROOT-CA sometimes perform slightly better than standard PMI-based methods and BERT. Experimental results on text classification for the BBCNews dataset show that ROOTROOT-CA and ROOT-CA outperform other methods under certain experimental settings.
In recent years, Document Question Answering (DocQA) tasks have placed greater demands on long-context processing and complex reasoning. However, current mainstream multimodal Retrieval-Augmented Generation (RAG) models predominantly rely on page-level visual retrieval. When processing high-density heterogeneous documents, these models are susceptible to “feature dilution” and are constrained by the “logical fragmentation” caused by physical pagination, which limits the precision of fine-grained evidence localization. To alleviate the aforementioned issues of granularity mismatch and semantic disconnection, this paper proposes MMRDoc: a document reasoning framework based on structural awareness and multi-granularity synergistic retrieval. First, MMRDoc bridges cross-layout and cross-modal semantic fragmentation through Semantic Integrity Reconstruction (SIR). Subsequently, it introduces a Synergistic Dual-path Retrieval mechanism to retrieve the global visual information of Top-K macro-pages and the fine-grained semantic features of micro-elements in parallel. To address the issues of information redundancy and logical fragmentation caused by multi-granularity retrieval, the model dynamically aggregates the heterogeneous candidate sets and subjects them to a Structural-aware Reranking module for unified evaluation. This achieves a logical transition from coarse-grained recall to fine-grained filtering. Experiments demonstrate that MMRDoc attains state-of-the-art (SOTA) performance on three major benchmarks: MMLongBench-Doc, PaperTab, and FetaTab, while maintaining strong robustness on the extreme-length LongDocURL benchmark. Compared to existing visual RAG baselines and multi-agent systems, MMRDoc enhances cross-page table understanding and long-range reasoning capabilities, providing an effective structured synergistic approach for multimodal long-document understanding.
Given the exponential increase of the information available online, we need effective and efficient methods for enabling the understanding and retrieval of key insights from complex data. Entity Summarization (ES) focuses on capturing the most relevant and representative information about an entity, facilitating its quick exploration and understanding. Previous entity summarization approaches struggled either to eliminate irrelevant information or to identify redundant content with an impact on the quality of the result. In this paper, we propose EntitySum, a novel approach that focuses on identify informative triples mainly based on entities’ topological and semantic importance as well as minimize redundancy. EntitySum, first filtering out noisy data, then selects the most relevant triples, considering enitites’ objects part in Knowledge Graph (KG) by using centrality measures combined with their properties’ frequency. Finally, to succeed the non-redundancy, we apply a minimize redundancy function that targeting the duplicate objects. Experiments conducted on the ESBM benchmark confirm that our method outperforms several baselines, demonstrating improved quality in entity summaries.
Sequential recommendation aims to predict future interactions by modeling historical user behavior, traditionally utilizing stacked Transformer architectures to capture interest evolution. However, pervasive data sparsity often triggers representation degradation, causing deep-layer features to homogenize, which impairs both expressive capability and optimization efficiency. Although data augmentation mitigates sparse signals, it frequently compromises semantic consistency across layers, undermining model robustness. Moreover, existing models rarely account for varying sparsity levels across datasets, which demand adaptive structural complexity. To address these challenges, we propose Distilled and Sparsity Adaptive Mamba for Sequential Recommendation (DiSAM4Rec). Leveraging Mamba’s linear complexity, we first design a Mamba-based Sparsity-Aware Dual-Branch (MSADB) module that dynamically switches between dense and sparse branches to optimize feature extraction across varying interaction densities. Based on this, we introduce a Layer-wise Semantic Alignment Self-Distillation (LSA-SD) mechanism to align intermediate features with high-level abstract semantic spaces, thereby mitigating representation degradation and ensuring deep-layer consistency. Extensive experiments on four real-world datasets demonstrate that DiSAM4Rec significantly outperforms competitive benchmarks, achieving average improvements of 4.12 https://github.com/HammerLau/DiSAM4Rec .
Community detection is essential for uncovering the functional organization of complex networks. While traditional methods often rely on edge density, motif-based approaches use higher-order structural patterns to identify communities. However, existing research frequently employs conventional motifs, such as triangles or 4-node cliques, or lacks validation against networks with ground-truth communities. This study addresses these limitations by systematically evaluating eight small motifs across both synthetic and real-world networks with known community structures. We propose a framework that transforms unweighted graphs into weighted representations by assigning weights to node pairs based on their co-occurrence frequency within specific graphlets, while also preserving information about the original edges, rather than creating a potentially sparse (hyper)network. Thus, graphlet adjacency captures the topological complexity of a node by accounting for both its direct edges and the local connectivity patterns of its neighbors; this higher-order information is vital for accurate community detection. Our results demonstrate that graphlet-based weighting significantly enhances community detection in networks. We find that no single "universal" motif optimizes performance across all real-world networks. Rather than favoring only dense, clique-based structures, our findings highlight that simpler motifs can also provide strong performance in networks. These results suggest that relying exclusively on cliques may overlook critical connectivity patterns, offering a new perspective on how higher-order structures define communities in networks.
Query expansion aims to address the term mismatch problem by adding relevant terms to the original query. Traditional approaches, particularly those based on relevance feedback, have shown significant improvements in information retrieval, but rely mainly on lexical information for selecting expansion terms. In contrast, contextual models such as BERT have proven effective in capturing semantic relationships between queries and documents, leading to improved retrieval performance. In this work, we propose an extension of the classical relevance model that incorporates BERT-based contextual information into query model estimation. Specifically, our approach leverages passage-level signals to better estimate the importance of expansion terms. The resulting model is combined with the original relevance model to form a hybrid query model for document retrieval. Experiments on four TREC collections demonstrate that our approach outperforms the baseline relevance model and traditional query expansion methods. It also surpasses two BERT-based query expansion models, CEQE and SRoc, across most evaluation metrics, while achieving competitive performance compared to the dense retrieval model ColBERT. These results highlight the effectiveness of the proposed contextualized hybrid expansion approach.
The rapid growth of high-resolution medical time-series data presents significant challenges for both efficient processing and clinical interpretability. We propose an interpretable data-processing architecture that transforms large-scale physiological data into compact, clinically meaningful representations. Our approach is built on a filter-centric framework in which simple, computationally efficient conditions are systematically constructed to capture salient patterns in patient trajectories. Filters are evaluated through an exhaustive and reproducible search procedure, enabling the identification of highly discriminative patterns without relying on opaque black-box models. To ensure reliability in clinical applications, we introduce a principled selection strategy that prioritizes filters achieving perfect precision while maximizing coverage, thereby identifying patient groups with highly consistent outcomes and providing transparent, trustworthy decision support. To support large-scale deployment, we develop a multi-resolution indexing and scoring mechanism that accommodates incremental updates as new data are collected, allowing continuous integration of longitudinal patient records. We demonstrate the effectiveness of the framework using the suppression ratio as an illustrative feature; however, the methodology is general and applicable to a wide range of physiological signals and time-series representations. Our results confirm that the proposed architecture delivers scalable, interpretable, and clinically meaningful analysis of complex medical time-series data.
Conversational Recommender Systems (CRSs) are designed to deeply understand user intent and provide personalized recommendations and responses through natural language interactions. However, existing methods primarily face two core challenges. First, the naive fusion of multi-source heterogeneous information tends to introduce redundant and conflicting signals into user representations, limiting the fine-grained characterization of user preferences. Second, factors such as colloquial expressions and noise within knowledge graphs significantly compromise the robustness of user representations, thereby affecting the stability of both recommendation and conversation tasks. To address these issues, this paper proposes a Multi-View Hypergraph Disentanglement and diffusion denoising framework (MVHD). Specifically, the proposed method first constructs three complementary hypergraph views–collaborative attributes, structural semantic, and similar user groups–to structurally characterize user preferences from distinct perspectives. A dynamic feature disentanglement mechanism is then employed to explicitly extract common and view-specific features, followed by personalized integration to construct a well-disentangled initial user representation. Furthermore, a condition-guided diffusion denoising strategy is designed. By leveraging collaborative attribute information as guidance, this strategy explicitly purifies noise within the initial user representation through forward noise injection and reverse denoising processes, ultimately yielding a more robust and refined user representation. Extensive experiments on two benchmark datasets demonstrate that MVHD significantly outperforms existing state-of-the-art methods in both recommendation and conversation subtasks.
The proliferation of reversal events on social media exacerbates fake news dissemination, undermining public trust and civil order. Such news often resurfaces across different periods, continuing to mislead the public even after debunking, posing a major challenge for detection. Existing studies primarily focus on label-specific correlations but overlook temporal shifts, reducing detection and generalization performance on unseen events. To address this, we introduce ReversalSV, a large-scale Chinese reversal multimodal news dataset covering diverse reversal events. Through macro- and micro-temporal analysis, we reveal unique temporal characteristics of such reversal event multimodal news. Additionally, we propose a Brain-inspired Memory-Replay Model (BMRM), which makes the first attempt to simulate human memory mechanisms for detecting fake news in reversal events. Inspired by memory replay theory, BMRM integrates semantic cognition and episodic memory for reversal fake news verification. Extensive experiments demonstrate BMRM’s superiority and establish a new paradigm for real-world multimodal fake news detection while providing a benchmark for evaluating models’ adaptability to evolving information landscapes.
Digital Twins (DTs) are increasingly integrating artificial intelligence (AI) to enhance predictive and adaptive capabilities, giving rise to AI-integrated Digital Twins (AI-DTs). While existing studies demonstrate the effectiveness of AI techniques within specific DT applications, they often lack an explicit system-level conceptualisation, leaving information flows, model interactions, and data transformations implicit. This limits transparency, explainability, and the informed use of AI-driven outputs. Meanwhile, information modelling approaches for conventional DTs offer limited support for representing computational models—including AI models—and their interactions with one another and with physical systems. This paper addresses these gaps by characterising information flows in AI-DTs, identifying key information modelling requirements, and proposing a generic semantic model that represents physical systems, computational models, data artefacts, and their interrelationships. By making data provenance, model dependencies, and transformation processes explicit, the proposed model supports transparency and facilitates explainability of AI-DT workflows. The feasibility of the model is demonstrated through extension and evaluation in a robot arm use case, illustrating its capability to represent and query these relationships in an integrated manner.
Query optimizers select efficient execution plans to minimize query latency and resource consumption in database systems, where cardinality estimators serve as their core component. While learned cardinality estimators have achieved significant improvements in relational databases by optimizing multi-table joins, vector databases face fundamentally different challenges with complex single-table predicate filtering and approximate nearest neighbor queries. To address these issues, we propose VecCard, a learned cardinality estimation model tailored for predicate filtering workloads in vector databases. Specifically, we design a Tree-LSTM-based architecture that accurately predicts the cardinalities of predicate filter queries through hierarchical semantic modeling. To enable flexible adaptations in dynamic environments without costly full retraining, we introduce an elastic weight consolidation (EWC) based incremental learning mechanism that preserves essential knowledge while learning new data and query patterns. Experimental results on the open-source vector database, DingoDB, demonstrate that VecCard yields more accurate cardinality estimates than several state-of-the-art baselines. It also significantly improves the performance of the generated query execution plans when integrated with DingoDB’s native optimizer.
Sequential recommendation aims to predict the next item a user is likely to engage with based on historical interactions. Traditional methods mainly focus on the order and identifiers of interacted items, often overlooking additional item-related information that could enrich user preference modeling. Many online platforms, however, provide textual reviews that reflect users’ opinions, experiences, and decision-making rationales. Leveraging both sequential behavioral patterns and review content can yield more comprehensive user representations and improve prediction. This study introduces RETSRec (Review-Enhanced Transformer-based Model for Sequential Recommendation), a novel recommendation architecture that integrates sequential behavioral signals with semantic insights derived from reviews. The model employs two encoding branches: 1) a bidirectional transformer to capture dependencies within the ordered interaction sequence, and 2) a BERT-like pre-trained language model followed by a hierarchical attention mechanism to extract high-value semantic features from reviews. By applying multi-head self-attention on aggregated reviews from different items, RETSRec allows modeling contextual dependencies among reviews, thus uncovering consistent behavioral patterns and latent user preferences. Experiments on real-world Amazon datasets demonstrate that RETSRec achieves significant improvements over state-of-the-art methods, underscoring the effectiveness of its design.
