The learned plan-selection optimizers, which generate diverse candidate plans via different traditional optimizers and then select the best expected performance plan using a learned value model, can achieve stable and relatively efficient performance by combining the advantages of conventional and learned methods. However, these eagerly-generated plans incur high optimization overhead, as they require multiple invocations of the native optimizer. In this paper, we propose MoEPlan, a method that learns a routing policy to select the top- k experts (different optimizers) by directly matching query embedding and expert embeddings, without needing to explicitly obtain execution plans from each expert in advance. MoEPlan also incorporates a virtual ideal expert, which is trained independently and guides the selection of the best plan of the top-k experts through learned plan similarities. Experimental studies demonstrate that MoEPlan, with only two plans generated, takes less inference time, while still producing more efficient plans than other learned plan-selection optimizers.

Multimodal Retrieval-Augmented Generation (RAG) with visual citation is crucial for ensuring the traceability and verifiability of MLLMs. However, current RAG and SFT-based methods struggle to achieve robust cross-modal reasoning, causing imprecise visual citations or decoupling between the citation and the generated answers. To address these limitations, we propose MCite-RL, a citation-enhanced agentic reinforcement learning framework designed for reliable multimodal RAG. MCite-RL introduces an Agentic Refinement module for visual citation that employs iterative retrieval, reasoning, and recursive cropping to progressively narrow the search space, transforming citation into a dynamic, evidence-driven reasoning process rather than a static step. Furthermore, we incorporate a Citation-enhanced Reward mechanism that integrates both process-level and outcome-level feedback within a reinforcement learning paradigm to jointly optimize answer accuracy and source traceability. Extensive experiments on benchmarks such as Wiki-VISA, FinRAGBench-V, and MMLongBench-Doc demonstrate that MCite-RL effectively achieves the joint optimization of citation precision and answer quality.

Learned query optimizers (LQOs) have shown remarkable progress in recent years. Despite achieving competitive performance compared with traditional methods, current LQOs struggle to retain stability and generate efficient plans when exposed to workloads with substantial diversity, presenting a major challenge in real-world deployment. Divo is a learned query optimizer designed to overcome performance degradation observed in training on diverse workloads. First, Divo proposes a template-evolving query generator to provide diverse queries for sufficient training. By modifying and combining query templates, the query generator synthesizes 3,000 informative new queries with fidelity to existing workloads. Second, Divo establishes a two-phase model training pipeline to enhance RL training with numerous plans collected in advance. We collected over 100,000 plans from public workloads and diverse generated queries to form static experiences, which are effectively leveraged on various training workload configurations to enhance Divo's plan generation through augmenting auxiliary components. Third, Divo proposes a diversity-aware loss function to learn from diversified input queries stably. By learning to predict probability distributions of latencies, Divo flexibly handles various queries and tolerates inconsistent ground truth latencies from training queries. We evaluate Divo's performance on PostgreSQL using a complicated mixed workload of JOB, DSB, Extended JOB, Stack, and generated queries, with three different training and testing workload configurations split by template. Experiments demonstrate Divo's advantage on total execution latency with a 1.3 & times; speedup relative to PostgreSQL, which is 14.9 & times; better than six existing LQOs on average.

Query optimization is essential for efficient SQL query execution in DBMS, and remains attractive over time due to the growth of data volumes and advances in hardware. Existing traditional optimizers struggle with the cumbersome hand-tuning required for complex workloads, and the learning-based methods face limitations in ensuring generalization. With the great success of Large Language Model (LLM) across diverse downstream tasks, this paper explores how LLMs can be incorporated to enhance the generalization of learned optimizers. Though promising, such an incorporation still presents challenges, mainly including high model inference latency, and the substantial fine-tuning cost and suboptimal performance due to inherent discrepancy between the token sequences in LLM and structured SQL execution plans with rich numerical features. In this paper, we focus on recurring queries in offline optimization to alleviate the issue of high inference latency, and propose LLM4Hint that leverages moderate-sized backbone LLMs to recommend query optimization hints. LLM4Hint achieves the goals through: (i) integrating a lightweight model to produce a soft prompt, which captures the data distribution in DBMS and the SQL predicates to provide sufficient optimization features while simultaneously reducing the context length fed to the LLM, (ii) devising a query rewriting strategy using a larger commercial LLM, so as to simplify SQL semantics for the backbone LLM and reduce fine-tuning costs, and (iii) introducing an explicit matching prompt to facilitate alignment between the LLM and the lightweight model, which can accelerate convergence of the combined model. Experiments show that LLM4Hint, by leveraging the LLM's stronger capability to understand the query statement, can outperform the state-of-the-art learned optimizers in terms of both effectiveness and generalization.

Hybrid queries, which combine vector nearest neighbor searches with scalar predicates, represent a fundamental challenge in managing vector databases. Existing methods often restrict the number of vector columns involved or the complexity of scalar predicates, thereby limiting their flexibility in handling diverse query patterns. Moreover, these approaches typically do not fully leverage the correlations between scalar and vector attributes, or the distributional patterns observed from query vector neighborhoods. To address these limitations, we introduce BoomHQ, a learning-based framework to boost multiple hybrid queries on vector DBMSs. First, BoomHQ models the correlation between vector and scalar attributes using an autoencoder-based architecture, which is also friendly to data updates. Second, BoomHQ captures prevailing query patterns, particularly using estimated selectivity of scalar predicates within the neighborhood of a query vector. Guided by these two key features, BoomHQ predicts the execution hints and rewrites the original query into an optimized version. Furthermore, we extend well-known benchmarks by introducing vector and scalar data with inherent correlations to better evaluate query execution. Experimental results demonstrate that for multiple hybrid queries at specified recall thresholds, our method achieves a 2x average and over 25x peak speedup compared to the state-of-the-art. Additionally, BoomHQ shows strong robustness against data updates and consistent optimization effectiveness across three representative vector database systems.

Most large enterprises build predefined data pipelines and execute them periodically to process operational data using SQL queries for various tasks. A key issue in minimizing the overall makespan of these pipelines is the efficient scheduling of varying concurrent queries within the pipelines. Existing tools mainly rely on simple heuristic rules due to the difficulty of expressing the complex features and mutual influences of queries. The latest reinforcement learning (RL) based methods have the potential to capture these patterns from feedback, but it is non-trivial to apply them directly due to the large scheduling space, high sampling cost, poor sample utilization, and limited generalization ability. Motivated by these challenges, we propose BQSched ^+ , a generalizable RL-based Scheduler for varying Batch concurrent Queries. Specifically, we design a grouped attention-based state representation that captures the complex query patterns efficiently by incorporating the interaction patterns of different query groups and supports varying batch queries. We also propose IQ-PPO-m, an auxiliary task-enhanced proximal policy optimization (PPO) algorithm with multi-task balancing, to fully exploit the rich signals of Individual Query completion in logs and enable the learned policy to adapt to varying query sets. Based on the RL framework above, we further introduce three optimization strategies, including adaptive masking to prune the action space, scheduling gain-based query clustering to deal with large query sets, and an incremental simulator to reduce sampling cost. Extensive experiments show that BQSched ^+ can significantly improve the efficiency and stability of batch query scheduling, while also achieving remarkable scalability and generalizability in both data and queries. For example, across all DBMSs and scales tested, BQSched ^+ reduces the overall makespan of batch queries on TPC-DS benchmark by an average of 34 ^+ is available at https://github.com/chxu2000/BQSched .

Large language models (LLMs) are increasingly relied upon for multi-turn dialogue to conduct complex tasks. However, existing benchmarks mainly evaluate LLMs as agents, overlooking their potential as independent systems to accomplish complex tasks. In addition, these benchmarks typically evaluate the planning and completion capabilities of the models individually, rather than simultaneously. To address these issues, we propose a new Dynamic Evaluation Framework for Multi-Turn task planning and completion (DEF-MT) to assess the ability of LLM to independently complete complex tasks in multi-turn scenarios. Our approach quantifies the model's planning capability by guiding it to generate planning and responses sequentially. Simultaneously, we use a dynamic approach to generate data that simulates the complex intents of real users. Finally, experiments conducted on 9 mainstream models using the Multiwoz 2.2 dataset, indicate that the existing models' sub-task planning capabilities hinder their ability to complete complex tasks, providing a meaningful reference for the future optimization direction of LLM.

Graph matching is a critical task with diverse real-world applications. Current cutting-edge methodologies incorporate GNN (Graph Neural Network) combined with incremental anchor refinement, calculating the matching similarity directly via node embeddings. However, the direct similarity computation based on aggregated embeddings from GNN may obscure the distinctiveness of nodes within a localized region. In addition, the possible wrongly added anchor pairs in the iterations and the lack of capturing the relationships to anchors may further affect the performance. In order to tackle these challenges, this paper proposes a method named DeepNM, which attempts to find node matching based on their neighbors' similarities. Specifically, DeepNM introduces a Sinkhorn-based similarity on a node's neighborhood's embeddings, which serves as both a training loss and a matching metric tailored to the graph matching problem. Additionally, we demonstrate that the Sinkhorn-based similarity, which relies on common neighbor statistics, is highly resilient to inaccurately identified anchor pairs within the context of incremental graph matching. Our comprehensive experiments on synthetic and real-world datasets demonstrate that DeepNM, compatible with the incremental graph matching paradigm, excels particularly well at matching graphs where common neighbors provide good matches. Applying the DeepNM pipeline to real social networks results in a 6% improvement, and applying the Sinkhorn similarity on knowledge graphs results in an average improvement of 1.7% over the best baseline.

Accurate cardinality estimation of substring queries, which are commonly expressed using the SQL LIKE predicate, is crucial for query optimization in database systems. While both rule-based methods and machine learning-based methods have been developed to optimize various aspects of cardinality estimation, their absence of error bounds may result in substantial estimation errors, leading to suboptimal execution plans. In this paper, we propose SSCard, a novel SubString Cardinality estimator that leverages a space-efficient FM-Index into flexible database applications. SSCard first extends the FM-Index to support multiple strings naturally, and then organizes the FM-index using a pruned suffix tree. The suffix tree structure enables precise cardinality estimation for short patterns and achieves high compression via a pushup operation, especially on a large alphabet with skewed character distributions. Furthermore, SSCard incorporates a spline interpolation method with an error bound to balance space usage and estimation accuracy. Additional innovations include a bidirectional estimation algorithm and incremental update strategies. Extensive experimental results in five real-life datasets show that SSCard outperforms both traditional methods and recent learning-based methods, which achieves an average reduction of 20% in the average q-error, 80% in the maximum q-error, and 50% in the construction time, compared with second-best approaches.

Macro placement, which involves optimizing the positions of modules, is a critical phase in modern integrated circuit design and significantly influences chip performance. The growing complexity of integrated circuits demands increasingly sophisticated placement solutions. Existing approaches have evolved along two primary paths (e.g., constructive and adjustment methods), but they face significant practical limitations that affect real-world chip design. Recent hybrid frameworks such as WireMask-EA have attempted to combine these strategies, but significant technical barriers still remain, including the computational overhead from separated layout adjustment and reconstruction that often require complete layout rebuilding, the inefficient exploration of design spaces due to random mutation operations, and the computational complexity of mask-based construction methods that limit scalability. To overcome these limitations, we introduce EGPlace, a novel evolutionary optimization framework that combines guided mutation strategies with efficient layout reconstruction. EGPlace introduces two key innovations: a greedy repositioning-guided mutation operator that systematically identifies and optimizes critical layout regions, and an efficient mask computation algorithm that accelerates layout evaluation. Our extensive evaluation using ISPD2005 and Ariane RISC-V CPU benchmarks demonstrate that EGPlace reduces wirelength by 10.8% and 9.3% compared to WireMask-EA and the state-of-the-art reinforcement learning-based constructive method EfficientPlace, respectively, while achieving speedups of 7.8x and 2.8x over these methods.

Table union search (TUS) represents a fundamental operation in data lakes to find tables unionable to the given one. Recent approaches to TUS mainly learn column representations for searching by introducing Pre-trained Language Models (PLMs), especially on columns with linguistic data. However, a significant amount of non-linguistic data, notably represented by domain-specific strings and numerical data in the data lake, are still under-explored in the existing methods. To address this issue, we propose LIFTus, an adaptive multi-aspect column representation for table unionable search, where aspect refers to a concept more flexible than data types, so that a single column can exhibit multiple aspects simultaneously. LIFTus aims at combining different aspects of a column (including both linguistic and non-linguistic aspects) to promote the effectiveness and generalization of TUS in a self-supervised manner. Specifically, besides employing PLMs to extract the linguistic aspects from an individual column, LIFTus trains a pattern encoder to learn possible character-level sequential patterns for the column, and builds a number encoder to capture numerical aspects of the column, including the distribution and magnitude features. LIFTus further utilizes a hierarchical cross-attention aided by aspect-relevant statistics to combine these aspects adaptively in producing the final column representations, which are indexed by vector retrieval techniques to achieve efficient search. Extensive experimental results demonstrate that LIFTus has outperformed the current state-of-the-art methods in terms of effectiveness, and achieved much better generalization capability to support unseen data.

The wide use of abbreviated column names (derived from English words or Chinese Pinyin) in database tables poses significant challenges for table-centric tasks in natural language processing and database management. Such a column name expansion task, referred to as the NameGuess task, has previously been addressed by fine-tuning Large Language Models (LLMs) on synthetically generated rule-based data. However, the current approaches yield suboptimal performance due to two fundamental limitations: 1) the rule-generated abbreviation data fails to reflect real-world distribution, and 2) the failure of LLMs to follow the rule-sensitive patterns in NameGuess persistently. For the data realism issue, we propose a novel approach that integrates a subsequence abbreviation generator trained on human-annotated data and collects non-subsequence abbreviations to improve the training set. For the rule violation issue, we propose a decoding system constrained on an automaton that represents the rules of abbreviation expansion. We extended the original English NameGuess test set to include non-subsequence and PinYin scenarios. Experimental results show that properly tuned 7/8B moderate-size LLMs with a refined decoding system can surpass the few-shot performance of state-of-the-art LLMs, such as the GPT-4 series. The code and data are presented in the supplementary material.

Most large enterprises build predefined data pipelines and execute them periodically to process operational data using SQL queries for various tasks. A key issue in minimizing the overall makespan of these pipelines is the efficient scheduling of concurrent queries within the pipelines. Existing tools mainly rely on simple heuristic rules due to the difficulty of expressing the complex features and mutual influences of queries. The latest reinforcement learning (RL) based methods have the potential to capture these patterns from feedback, but it is non-trivial to apply them directly due to the large scheduling space, high sampling cost, and poor sample utilization. Motivated by these challenges, we propose BQSched, a non-intrusive Scheduler for Batch concurrent Queries via reinforcement learning. Specifically, BQSched designs an attention-based state representation to capture the complex query patterns, and proposes IQ-PPO, an auxiliary task-enhanced proximal policy optimization (PPO) algorithm, to fully exploit the rich signals of Individual Query completion in logs. Based on the RL framework above, BQSched further introduces three optimization strategies, including adaptive masking to prune the action space, scheduling gain-based query clustering to deal with large query sets, and an incremental simulator to reduce sampling cost. To our knowledge, BQSched is the first non-intrusive batch query scheduler via RL. Extensive experiments show that BQSched can significantly improve the efficiency and stability of batch query scheduling, while also achieving remarkable scalability and adaptability in both data and queries. For example, across all DBMSs and scales tested, BQSched reduces the overall makespan of batch queries on TPC-DS benchmark by an average of 34 the commonly used heuristic strategy and the adapted RL-based scheduler, respectively.

The detection of malicious social bots has become a crucial task, as bots can be easily deployed and manipulated to spread disinformation, promote conspiracy messages, and more. Most existing approaches utilize graph neural networks (GNNs)to capture both user profle and structural features,achieving promising progress. However, they still face limitations including the expensive training on large underlying graph, the performance degration when similar neighborhood patterns' assumption preferred by GNNs is not satisfied, and the dynamic features of bots in a highly adversarial context. Motivated by these limitations, this paper proposes a method named BSG4Bot with an intuition that GNNs training on Biased SubGraphs can improve both performance and time/space efficiency in bot detection. Specifically, BSG4Bot first pre-trains a classifier on node features efficiently to define the node similarities, and constructs biased subgraphs by combining the similarities computed by the pre-trained classifier and the node importances computed by Personalized PageRank (PPR scores). BSG4Bot then introduces a heterogeneous GNN over the constructed subgraphs to detect bots effectively and efficiently. The relatively stable features, including the content category and temporal activity features, are explored and incorporated into BSG4Bot after preliminary verification on sample data. The extensive experimental studies show that BSG4Bot outperforms the state-of-the-art bot detection methods, while only needing nearly 1/5 training time.

Denoising Diffusion Probabilistic Models (DDPMs) have recently demonstrated exceptional performance in generating high-quality data. In this work, we propose CommunityDF, a novel framework that applies DDPMs to the community search problem, which involves identifying subgraphs containing nodes closely related to a given query node. However, three key challenges arise in this context: (I) learning effective node representations from limited examples, (II) discretizing continuous node representations into community members, and (III) reducing the number of diffusion steps without sacrificing performance. To tackle these, CommunityDF introduces several innovations. First, we focus on subgraphs around the query node to reduce interference from unrelated nodes, improve scalability. We then employ a contrastive learning approach, treating node states at different diffusion steps as positive examples and designing various negative sampling strategies to learn high-quality node representations from limited examples. Second, we propose a dynamic thresholding mechanism that effectively converts continuous representations into community members. Finally, we reduce the number of diffusion steps by leveraging the rough communities to initialize the process with rough community structures, which accelerates convergence while maintaining high accuracy. Extensive experiments on seven real-world datasets demonstrate that CommunityDF outperforms existing methods by 16%-47%, establishing it as a state-of-the-art solution for community search. The source code is available at https://github.com/JiazunChen/CommunityDF.

Social bots are increasingly polluting online platforms by spreading misinformation and engaging in coordinated manipulation, posing severe threats to cybersecurity. Graph Neural Networks (GNNs) have become mainstream for social bot detection due to their ability to integrate structural and attribute features, with spectral-based approaches demonstrating particular efficacy due to discriminative patterns in the spectral domain. However, current spectral GNN methods face two limitations: (1) their broad-spectrum fitting mechanisms degrade the focus on bot-specific spectral features, and (2) certain domain knowledge valuable for bot detection, e.g., low homophily correlates with high-frequency features, has not been fully incorporated into existing methods. To address these challenges, we propose HW-GNN, a novel homophily-aware graph spectral network with Gaussian window constraints. Our framework introduces two key innovations: (i) a Gaussian-window constrained spectral network that employs learnable Gaussian windows to highlight bot-related spectral features, and (ii) a homophily-aware adaptation mechanism that injects domain knowledge between homophily ratios and frequency features into the Gaussian window optimization process. Through extensive experimentation on multiple benchmark datasets, we demonstrate that HW-GNN achieves state-of-the-art bot detection performance, outperforming existing methods with an average improvement of 4.3

Retrieval-Augmented Generation (RAG) plays a vital role in the financial domain, powering applications such as real-time market analysis, trend forecasting, and interest rate computation. However, most existing RAG research in finance focuses predominantly on textual data, overlooking the rich visual content in financial documents, resulting in the loss of key analytical insights. To bridge this gap, we present FinRAGBench-V, a comprehensive visual RAG benchmark tailored for finance which effectively integrates multimodal data and provides visual citation to ensure traceability. It includes a bilingual retrieval corpus with 60,780 Chinese and 51,219 English pages, along with a high-quality, human-annotated question-answering (QA) dataset spanning heterogeneous data types and seven question categories. Moreover, we introduce RGenCite, an RAG baseline that seamlessly integrates visual citation with generation. Furthermore, we propose an automatic citation evaluation method to systematically assess the visual citation capabilities of Multimodal Large Language Models (MLLMs). Extensive experiments on RGenCite underscore the challenging nature of FinRAGBench-V, providing valuable insights for the development of multimodal RAG systems in finance.

This paper presents the db3 team's winning solution for the Meta CRAG-MM Challenge 2025 at KDD Cup'25. Addressing the challenge's unique multi-modal, multi-turn question answering benchmark (CRAG-MM), we developed a comprehensive framework that integrates tailored retrieval pipelines for different tasks with a unified LLM-tuning approach for hallucination control. Our solution features (1) domain-specific retrieval pipelines handling image-indexed knowledge graphs, web sources, and multi-turn conversations; and (2) advanced refusal training using SFT, DPO, and RL. The system achieved 2nd place in Task 1, 2nd place in Task 2, and 1st place in Task 3, securing the grand prize for excellence in ego-centric queries through superior handling of first-person perspective challenges.

Large language models (LLMs) excel in question-answering (QA) tasks, and retrieval-augmented generation (RAG) enhances their precision by incorporating external evidence from diverse sources like web pages, databases, and knowledge graphs. However, current RAG methods rely on agent-specific strategies for individual data sources, posing challenges low-resource or black-box environments and complicates operations when evidence is fragmented across sources. To address these limitations, we propose ER-RAG, a framework that unifies evidence integration across heterogeneous data sources using the Entity-Relationship (ER) model. ER-RAG standardizes entity retrieval and relationship querying through ER-based APIs with GET and JOIN operations. It employs a two-stage generation process: first, a preference optimization module selects optimal sources; second, another module constructs API chains based on source schemas. This unified approach allows efficient fine-tuning and seamless integration across diverse data sources. ER-RAG demonstrated its effectiveness by winning all three tracks of the 2024 KDDCup CRAG Challenge, achieving performance on par with commercial RAG pipelines using an 8B LLM backbone. It outperformed hybrid competitors by 3.1 by 5.5X.

This paper describes the winning solutions of all tasks in Meta KDD Cup 24 from db3 team. The challenge is to build a RAG system from web sources and knowledge graphs. We are given multiple sources for each query to help us answer the question. The CRAG challenge involves three tasks: (1) condensing information from web pages into accurate answers, (2) integrating structured data from mock knowledge graphs, and (3) selecting and integrating critical data from extensive web pages and APIs to reflect real-world retrieval challenges. Our solution for Task #1 is a framework of web or open-data retrieval and answering. The large language model (LLM) is tuned for better RAG performance and less hallucination. Task #2 and Task #3 solutions are based on a regularized API set for domain questions and the API generation method using tuned LLM. Our knowledge graph API interface extracts directly relevant information to help LLMs answer correctly. Our solution achieves 1st place on all three tasks, achieving a score of 28.4