In millimeter-wave ad hoc networks, the challenges of directional neighbor discovery (ND), combined with the emerging integration of reinforcement learning (RL), necessitate algorithms that are both efficient and RL-compatible under practical constraints. To address this, we propose the Constrained Random Algorithm (CoRA), a novel directional ND scheme achieving high discovery efficiency, robustness to node density variations, broad compatibility with enhancement techniques, and strong support for RL-based optimization. Specifically, CoRA constructs constrained beam direction sets using minimal constraint units—spatially symmetric direction pairs—and executes ND by sequentially scanning these sets with randomized beam selection. We analyze CoRA and show that it provides a unified analytical formulation that includes Pure Random Algorithm (PRA) and Scan-Based Algorithm (SBA) as limiting cases, achieving up to 80% lower time-slot overhead than PRA in sparse networks and up to 60% lower than SBA in dense networks, with strong performance even under minimal hardware and without enhancement techniques. To illustrate compatibility, we propose two variants: GCoRA employing gossip-based neighbor information exchange and SCoRA with a selective reply mechanism, both showing additional performance gains. Furthermore, we propose QCoRA by integrating Q-learning and benchmarking it against QPRA. QCoRA achieves faster convergence, higher cumulative rewards, and enhanced exploration. Finally, we extend QCoRA to deep RL by proposing DDQCoRA, which leverages Double DQN to further reduce discovery latency. Extensive simulations confirm the effectiveness and adaptability of the proposed CoRA family.

Schema matching is pivotal in data integration, identifying semantic equivalences between attributes in structured data tables. Existing methods focus on schema-level or instance-level matching, excelling only in specific scenarios. This underscores the need for a unified approach adaptable to varying data availability conditions. This paper presents a unified schema matching framework based on large language models (LLMs) that effectively addresses schema matching tasks across three distinct scenarios: a Scenario with only MetaData (SMD), a Scenario with a Small amount of instance Data (SSD), and a Scenario with a Large amount of instance Data (SLD).The framework employs a two-step approach, integrating feature similarity computation with LLM prompt engineering. The first step involves comprehensive feature extraction and similarity analysis to efficiently filter potential matches, followed by scenario-specific LLM prompting for precise semantic matching. Our approach integrates textual, structural, and semantic features with LLM-based analysis using LLaMA3.1-8b as its base model. Extensive experiments on WikiData and MIMIC-III datasets validate the framework’s effectiveness, achieving an F1 score of 92.5 and precision of 95.8 in the SLD scenario, while maintaining robust performance in scenarios with limited data. Ablation studies confirm that this two-step approach enhances both efficiency and accuracy, validating the framework’s capability to address diverse schema matching requirements within a unified solution.