Structure-based graph rewiring, a graph augmentation approach, modifies the graph structure for enhancing the graph connectivity while keeping the graph intrinsic semantics, improving node embeddings in heterogeneous graphs. To augment the graph connectivity while avoiding introducing more noise into the graph, this paper proposes a novel model EulerRewire for heterogeneous graph by identifying all pairs of nodes with odd degree and adding edges between these node pairs while the graph spectral gap is utilized to guide edge selection, adding edges as few and enhancing the graph connectivity as much as possible. Then, a dual-view encoder based on GCN and Transformer to jointly learn node embeddings in the rewired Eulerian graph due to their abilities in capturing local structure and high-order dependency, in which a circuit sequence is generated for each node in the rewired Eulerian graph, better capturing high-order dependency between nodes. Extensive experiments for three downstream tasks including node classification, node clustering and link prediction demonstrate the proposed EulerRewire is superior to state-of-the-art baselines in most datasets.
Real-time agentic bidders are autonomous systems that participate in programmatic advertising auctions on behalf of advertisers, making instantaneous decisions on how much to bid on ad impressions while optimizing advertiser-defined objectives such as cost-per-acquisition (CPA). These agents operate under strict budget constraints in a dynamic, multi-agent environment where auctions repeat with constantly evolving bidding dynamics. In this paper, we propose a novel budget-constrained, CPA-aware agentic bidder that learns an adaptive bidding strategy to maximize the acquisition of high-quality users while ensuring effective budget utilization under sparse and delayed feedback. At the core of our approach is a linear-time approximation algorithm that formulates bid selection under a budget constraint as a greedy knapsack optimization problem. The algorithm scales to billions of auctions per day and provides a theoretical bound on the optimality gap, yielding provable near-optimal performance. To maintain robustness in production environments, the agent incorporates real-time feedback to pace budget spend throughout the day and adapt to shifting market conditions. Without explicitly modeling competing agents, the system learns a pragmatic near-optimal bidding strategy that balances campaign performance and budget utilization. Empirical evaluation on a large-scale demand-side platform (DSP) demonstrates that the proposed agent achieves approximately 92% of the theoretical optimum in acquisition efficiency under budget constraints. Our contributions include: (1) a scalable, near-optimal knapsack-based approximation algorithm with provable bounds; (2) a real-time adaptive bidding agent that jointly addresses budget pacing and CPA optimization under sparse and delayed feedback; and (3) extensive large-scale production validation in repeated auction environments without dominant bidding strategies.
We benchmark large language models for entity matching (EM) using the Compound Entity Matching (ComEM) framework, evaluating 46 model configurations across 8 datasets under a fixed protocol. Beyond accuracy, we quantify API-equivalent token cost and analyze the impact of configurable reasoning/thinking modes. We find that open-weight models can rival proprietary APIs: GPT-OSS:120b attains 92.10 F1 while costing ∼ 20 × less than GPT-5(high). Reasoning controls have opposite effects across families—low is optimal for GPT-5, whereas high improves open-weight models—and Claude shows the largest sensitivity (Sonnet 4.5: +10.13 F1 with thinking; Opus 4.5: slight degradation at 2.6 × cost). Finally, several ultra-budget configurations under 0.10 reach up to 89 F1, enabling inexpensive deployment. Our study offers practical model-selection guidance for EM across diverse budget and infrastructure constraints.
Sequential recommendation aims to model the dynamic preferences of users from historical behavior sequences for accurate item prediction. State space models (SSM) such as Mamba have demonstrated strong capability in long-sequence processing with linear complexity. However, in sequential recommendation, Mamba exhibits limited sensitivity to high frequency and transient local dynamics, suffers from insufficient nonlinear expressivity, and fails to capture the oscillatory nature present in complex user preference evolution. To address this limitation, we introduce Osc-Mamba, an effective framework that enhances Mamba for recommendation with a chaotic oscillatory mechanism (Lee-oscillator). The oscillator introduces frequency-sensitive nonlinear responses that adaptively regulate hidden-state transitions, improving the model’s feature representation ability. We design three strategies to integrate oscillatory dynamics with Mamba. Experimental results demonstrate that our approach outperforms competitive baselines in three sequential recommendation datasets.
The rapid progress of Large Language Models (LLMs) and Knowledge Graphs (KGs) has created new opportunities for building intelligent systems that are simultaneously expressive, grounded, and capable of complex reasoning. However, each paradigm has distinct limitations: LLMs struggle with factual consistency, long term memory, and controllability, while KGs lack the generative fluency and adaptive reasoning required for open ended tasks. This survey provides the first comprehensive study of how these two paradigms are increasingly unified through the emerging paradigm of agentic AI systems composed of autonomous agents. The survey introduces agentic AI as a conceptual and architectural bridge that unifies symbolic and neural intelligence, charting a path toward more adaptive LLM–KG agentic ecosystems. To the best of our knowledge, this is the first survey to explore the synergy between knowledge graphs and LLMs from an Agentic AI perspective.
Machine learning models such as tree ensembles and gradient boosting are widely used in public-health analytics due to greater interpretability than deep learning. However, these models assume county independence, which fails for regional health outcomes exhibiting strong spatial dependence, leading to biased predictions that do not reflect how risk varies across regions. This work addresses this limitation with HS-GPBoost, a spatial-feature learning framework based on Gaussian Process Boosting that combines nonlinear feature effects with a spatial Gaussian Process, capturing unobserved regional structure while providing an interpretable decomposition of feature-driven and geographic variation. We benchmark the framework against linear, tree-based, spatial autoregressive, and Gaussian process baselines across ten county-level health outcomes from the 2024 County Health Rankings under both random and spatial cross-validation, and include a held-out Florida case study to examine spatial prediction and uncertainty behavior. Results confirm that modeling spatial dependence directly is essential for accurate and interpretable regional health prediction.
As LLMs continue to evolve, their ability across domains has become a critical challenge. In this paper, we present a unified evaluation protocol for assessing domain-specific fidelity, grounded in a definition of fidelity hallucinations. We categorize it into two subtypes: aprioristic hallucination, which reflects a model’s intrinsic knowledge, and inferential hallucination, which arises from the model’s reasoning based on supplementary context. Using a diverse collection of academic corpora, we construct and rigorously validate a QA dataset exploring both types of fidelity hallucinations across domains, leveraging several LLMs with over 70 billion parameters. We then conduct our evaluation on smaller, commercially deployed models and uncover performance disparities across domains even within the same model. Through comprehensive quantitative analyses, we pinpoint particular domains where hallucinations tend to be more pronounced, along with the conditions that contribute to their occurrence. Furthermore, we demonstrate that strategic variations in decoding methods can significantly impact fidelity hallucinations. These findings highlight the importance of domain-aware evaluation protocols and offer practical guidance for the effective deployment of models in real-world applications.
Time-decay heavy hitters are items whose recency-weighted frequencies dominate a data stream and are central to applications such as data-warehouse analytics, anomaly detection and web query log analysis. Existing sketch-based solutions adopt recent-hotness-unaware eviction policies, which, under tight memory budgets, lead to substantial accuracy degradation due to noise from hash collisions and the premature eviction of recent hot items. We present RHA-Sketch, a compact, high-throughput sketch with a recent-hotness-aware eviction policy. It partitions items into different levels of recent hotness and applies differentiated decay strategies to each class. We evaluate RHA-Sketch on real traces. Experimental results show that RHA-Sketch achieves high F1-score under tight memory budgets (e.g., 16 KB ). Moreover, it outperforms state-of-the-art schemes by reducing ARE by 61.9% and improving throughput by 57% . All related implementations are released as open source in a GitHub repository [11].
This study evaluates scalable, unsupervised methods for detecting malicious online content by benchmarking transformer-based architectures across diverse linguistic contexts. We demonstrate that the semantic resolution of the embedding layer is the critical determinant of performance. Our results identify that large language models coupled with manifold learning achieve superior anomaly separation in knowledge-intensive domains, significantly outperforming traditional BERT-based pipelines. Additionally, we reveal a topological dichotomy in detection strategies: Contrastive autoencoders offer robust stability in structured environments, whereas few-shot deviation learning (FATE) is essential for high-entropy, dynamic contexts such as politics and sports. These findings propose a shift toward context-aware architectures capable of adapting to the complex semantic landscape of modern web content.
Bayesian Dynamic Linear Models provide coherent probabilistic forecasting, yet their sequential learning can adapt too slowly after abrupt policy-driven regime shifts, producing transient bias precisely when accurate predictions are most valuable. We propose FCI-Retrieval-DLM, an extension that retrieves historically similar macro-financial episodes via the Financial Conditions Impulse on Growth (FCI-G) index and applies a reliability-gated correction to the baseline predictive mean within event windows. A Shuffled-Retrieval Control, which preserves the full intervention pipeline while destroying semantic alignment, serves as a placebo-style negative control to isolate the contribution of context-guided matching. In walk-forward NASDAQ forecasting, FCI-Retrieval-DLM yields modest directional improvements over Base-DLM in both point and distributional accuracy, with gains concentrated in onset-centered post-event windows. However, differences relative to the shuffled control remain small under the current configuration, indicating that evidence for a distinct semantic retrieval advantage is not yet conclusive.
University admission advisory is a high-stakes decision-making process in which applicants must interpret and reconcile heterogeneous institutional regulations, eligibility criteria, tuition policies, scholarships, deadlines, and procedures across multiple universities. While recent advances in Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) enable document-grounded question answering, existing QA benchmarks are largely built on homogeneous corpora and seldom capture the cross-institution and regulation-driven multihop reasoning required in real admission scenarios, especially for Vietnamese. We present VixSTORY (Vietnamese Cross-Institution University Admission Advisory QA), a benchmark designed to evaluate hierarchical and graph-structured reasoning under both intra-institution and cross-institution settings. VixSTORY is constructed via a controlled semi-automatic pipeline that combines document segmentation, embedding-based topic clustering within institutions, cross-institution topic alignment, and LLM-assisted question generation with strict post-validation for faithfulness. We further provide a comprehensive evaluation of representative retrieval and reasoning paradigms, from lexical/dense baselines to iterative multihop and graph-based RAG methods. Results show substantial gaps between flat retrieval and structured approaches, with graph-based methods achieving markedly stronger performance and robustness as hop depth increases, highlighting the importance of explicit structure for realistic regulation-driven advisory systems. For reproducibility and future research, VixSTORY is publicly released at https://huggingface.co/datasets/ura-hcmut/VixSTORY .
Financial risk models provide critical quantitative insights, yet translating complex risk data into actionable narratives remains a challenge. While LLMs offer powerful text generation capabilities, they struggle with numerical reasoning and are prone to hallucinations which is a critical flaw in financial compliance environment. This paper proposes TSRMTGen, a framework that transforms risk model data into accurate, coherent narratives. Our approach integrates expert knowledge through “Hint” injection and utilizes a two-step generation process to ensure factual consistency. We demonstrate the effectiveness of TSRMTGen through a rigorous 3-layer human-in-the-loop evaluation, achieving a 0
This study investigates how to align knowledge graph with user preferences to enhance collaborative filtering recommendations. Existing knowledge-enhanced methods often rely on single-scenario neighbor aggregation, limiting their ability to capture diverse and uncertain user intents in real-world settings. To address these limitations, we propose MS-KIM, a Multi-scenario Knowledge-enhanced Latent Intent Modeling framework. At its core, MS-KIM formulates user intents as latent variables inferred through variational inference based on interactions and scenario-aware knowledge signals. Specifically, we first design a scenario-aware knowledge aggregation module to extract informative item representations under different scenarios. Building upon these signals, we leverage variational inference to jointly model user intents across multiple scenarios. Moreover, since scenario labels are often incomplete, we introduce a scenario-aware self-supervised task that promotes representation separation across scenarios by leveraging augmented subgraph-based user representations as guidance. Extensive experiments on three real-world datasets demonstrate that our model outperforms strong baselines.
Social media provides valuable insights into public sentiment, but analyzing sentiment accurately remains challenging due to the widespread presence of multi-level comment structures (MLCS) in social media texts. Existing research lacks dedicated datasets and systematic evaluation metrics for MLCS-based sentiment analysis tasks. This study proposes WeiboDRP, a large-scale dataset explicitly designed for Chinese sentiment analysis with MLCS annotation. Benchmark evaluations using dictionary methods, fine-tuned pre-trained models, and zero-shot large language models on this dataset reveal that DeepSeek-R1 achieves the best overall performance. Additionally, this study proposes a novel evaluation metric called the Hierarchical Consistency Score (HCS), which effectively measures models’ ability to capture and utilize hierarchical contextual information in MLCS. The effectiveness of HCS is verified through ablation experiments. Its cross-domain generalizability is further demonstrated on a publicly available benchmark dataset. These findings highlight the importance of effectively understanding and leveraging the contextual information provided by MLCS to improve classification performance in social media-based sentiment analysis tasks.
Large language models (LLMs) in medical question answering (QA) have recently revealed fundamental limitations that hinder their clinical deployment, including an inability to perform causal reasoning beyond pattern matching, the generation of hallucinated responses lacking biomedical grounding, and the failure to quantify reasoning confidence. To address these challenges, we propose CauMed, a causal graph-based framework for clinical reasoning in the medical QA task. CauMed integrates causal interference, biomedical fact validation, and patient-context adaptation. Our model first extracts clinically relevant facts and candidate causal relations from the input questions using a fine-tuned LLaMA-3 model. These relations are then validated via a BioBERT classifier trained in PrimeKG, forming a biomedical causal graph that supports structured inference. To select the final answer, CauMed applies a weight scoring function that combines causal graph support and semantic alignment, ensuring that responses are medically valid, clinically actionable, and contextually coherent. Our experimental results demonstrate that CauMed consistently outperforms baseline methods on the MedQA-US, MedMCQA, and MMLU-Med benchmarks, providing clinically grounded, highly reliable, and trustworthy diagnostic answers.
We present a four-layer architecture for automated journalism that addresses unbounded context growth, agent coordination complexity, and quality assurance at scale. The system comprises Layer 1 (Observability), Layer 2 (Specialized Agents), Layer 3 (Event-Driven Orchestration with Listener-Aware Communication), and Layer 4 (Hybrid Context Management with RAG). The architecture is implemented using Deep Agents, a LangGraph-based harness employing asynchronous message queues, agent pooling, domain-specific skills loaded via progressive disclosure, and remote MCP servers for external tool integration. The hybrid context strategy combines semantic compression (avg 56
Large language models (LLMs) are typically instruction tuned to improve task following, but this stage also enables back-door attacks. Prior work used Virtual Prompt Injection (VPI) for jailbreak/control steering; our contribution is to show VPI can serve as a bias-injection channel, embedding societal bias during instruction tuning. Poisoning as little as 2
The integration of Artificial Intelligence (AI) into Business Intelligence (BI) is vital as it enables businesses to process vast amounts of information quickly and uncover patterns and trends that might otherwise go unnoticed. In this short paper, we propose a five-phase AI adoption in the context of BI drawing from our experience with numerous SMEs in Singapore and internationally. This progression begins with AI for BI 1.0, where businesses rely on outsourced AI services, offering accessible but limited AI capabilities. As organizations seek greater control, AI for BI 2.0 emerges, emphasizing in-house AI development with a focus on interpretability, making AI models more understandable to users. AI for BI 3.0 advances this by adding explainability, clarifying the decision-making processes of AI systems. The transition to AI for BI 4.0 introduces trust as a crucial factor, ensuring that AI models are not only useful and robust but also faithful to the data they analyze. Finally, AI for BI 5.0, integrates human-centered design, enhancing decision-making through continuous feedback and adaptation. This phased evolution reflects the growing demands for transparency, trust, and collaboration in AI-driven BI, ultimately transforming AI from a tool into a trusted partner in business strategy.
Acceptance test automation remains a persistent bottleneck in software engineering, as translating natural-language requirements into executable specifications demands extensive expertise and manual effort. While large language models (LLMs) show promise in code generation, their integration into Behavior-Driven Development (BDD) pipelines—particularly for Gherkin test authoring—remains underexplored. We introduce CMAS4G2, a Cooperative Multi-Agent System for Gherkin Generation built on Microsoft’s AutoGen Swarm framework. CMAS4G2 coordinates specialized Validator, Quality-Assurance, and Supervisor agents to automate requirement validation and acceptance-test synthesis. Generated artifacts are evaluated at scale using the LLM-as-a-Judge (LAJ) framework [5], which provides rubric-driven, automated coverage assessment aligned with expert human judgments. Experiments on a 200-ticket benchmark derived from the Kill Bill platform show that open-weight models deployed locally on consumer hardware approach proprietary performance: GPT-OSS 20B (medium reasoning, our best-performing open-weight configuration by overall success) attains 87.6
