Hai Lan

Recent advances in large language models (LLMs) have motivated a new class of semantic data processing systems (SDPs) that integrate semantic reasoning into data processing over heterogeneous data lakes. Compared with traditional database systems, SDPs differ in three fundamental ways: (1) they operate over heterogeneous data lakes rather than curated relational tables; (2) they extend relational operators with LLM-powered semantic operators; (3) they increasingly take open-ended natural-language queries instead of precisely specified SQL. This survey provides a unified view of LLM-powered SDPs, including query analysis, semantic operator design, query optimization and execution, and relevant benchmarks. For each stage, we summarize representative techniques, their underlying motivations and design trade-offs. In addition, we further highlight major open challenges and future directions for this field.

Managing the rapidly growing scholarly corpus poses significant challenges in representation, reasoning, and efficient analysis. An ideal system should unify structured knowledge management, agentic planning, and interpretable execution to support diverse scholarly queries – from retrieval to knowledge discovery and generation – at scale. Unfortunately, existing RAG and document analytics systems fail to achieve all query types simultaneously. To this end, we propose AgenticScholar, an agentic scholarly data management system that integrates a structure-aware knowledge representation layer, an LLM-centric hybrid query planning layer, and a unified execution layer with composable operators. AgenticScholar autonomously translates natural language queries into executable DAG plans, enabling end-to-end reasoning over multi-modal scholarly data. Extensive experiments demonstrate that AgenticScholar significantly outperforms existing systems in effectiveness, efficiency, and interpretability, offering a practical foundation for future research on agentic scholarly data management.

Modern analytical workloads, particularly those powering business intelligence dashboards, are dominated by parameterized queries, where a small number of query templates are executed repeatedly with varying predicate values. Classical cost-based optimizers frequently fail to produce efficient execution plans in this setting due to inaccurate cardinality estimation and unstable plan choices. Although numerous learning-based optimizers have been proposed, most are designed for general, ad-hoc query optimization and overlook the repetitive structure of parameterized workloads. As a result, they often incur high training or serving overheads or require substantial modifications to existing database systems, limiting their practical adoption. We propose PLARQ, a practical learned optimizer tailored for Parameterized Query Optimization (PQO). We begin by analyzing existing approaches through a unified two-stage framework: (1) Plan Candidate Generation (PCG), which forms a set of plausible execution plans, and (2) Plan Ranking (PR), which selects the most promising one. This abstraction captures the core design principles of prior work and provides a foundation for systematic comparison. Building on this framework, PLARQ employs a similarity-based PCG approach to retrieve a compact, high-quality set of plan candidates from a precomputed plan pool, and a list-wise, attention-based ranking model to effectively identify the optimal plan among them. PLARQ integrates seamlessly with PostgreSQL without modifying the optimizer internals. Extensive experiments across five benchmarks demonstrate that PLARQ improves end-to-end query performance, achieving speedups of up to 420.31x over PostgreSQL and up to 2x over existing learned methods.

Missing values in tabular data lakes can severely impact data analysis and diminish the performance in downstream applications. We highlight that a robust imputation strategy should properly take three aspects of variety into consideration: source of imputed value, the types of tables involved, and the data types of the missing value. Existing imputation methods rely on estimation-based approaches (using a model trained on data from the same table to estimate missing values) or search-based approaches (retrieving values from other tables). Unfortunately, none of these approaches effectively incorporate all three aspects of variety. To address this gap, we propose , a novel framework that uses a C ombination of E stimation-based and S earch-based methods for missing value I mputation in D ata lakes. contains three core modules: (1) the , which efficiently discovers candidate values from tables by exploiting the contextual information; (2) the , which introduces an influence function and a sampling-based exploration strategy to yield accurate estimated values; (3) the , which determines the most suitable method based on table-level and column-level statistics. Extensive experiments conducted on three data lakes demonstrate that effectively and efficiently addresses the missing value problem.

Reinforcement learning has recently been used to enhance index structures, giving rise to reinforcement learning-enhanced spatial indices (RLESIs) that aim to improve query efficiency during index construction. However, their practical benefits remain unclear due to the lack of unified implementations and comprehensive evaluations, especially in disk-based settings. We present the first modular and extensible benchmark for RLESIs. Built on top of an existing spatial index library, our framework decouples index training from building, supports parameter tuning, and enables consistent comparison with traditional, advanced, and learned spatial indices. We evaluate 12 representative spatial indices across six datasets and diverse workloads, including point, range, kNN, spatial join, and mixed read/write queries. Using latency, I/O, and index statistics as metrics, we find that while RLESIs can reduce query latency with tuning, they consistently underperform learned spatial indices and advanced variants in both query efficiency and index build cost. These findings highlight that although RLESIs offer promising architectural compatibility, their high tuning costs and limited generalization hinder practical adoption.

Although many updatable learned indexes have been proposed in recent years, whether they can outperform traditional approaches on disk remains unknown. In this study, we revisit and implement four state-of-the-art updatable learned indexes on disk, and compare them against the B+-tree under a wide range of settings. Through our evaluation, we make some key observations: 1) Overall, the B+-tree performs well across a range of workload types and datasets. 2) A learned index could outperform B+-tree or other learned indexes on disk for a specific workload. For example, PGM achieves the best performance in write-only workloads while LIPP significantly outperforms others in lookup-only workloads. We further conduct a detailed performance analysis to reveal the strengths and weaknesses of these learned indexes on disk. Moreover, we summarize the observed common shortcomings in five categories and propose four design principles to guide future design of on-disk, updatable learned indexes: (1) reducing the index's tree height, (2) better data structures to lower operation overheads, (3) improving the efficiency of scan operations, and (4) more efficient storage layout.

This paper introduces a kind of immune genetic algorithm which regards objective function as antigen, solution as antibody and updates the population using evolutionary strategy. After economic emission load dispatch which belongs to multi-objective constrained optimization problems is discussed, main processes of immune genetic algorithm to solve this matter is given. Through tests a power system model with five coal-burning generating units, feasibility and validity of this algorithm is proved. And by comparing with genetic algorithm and Hopfield neural network, optimization and quick constringency of this algorithm to solve similar problems are proved