Compiling Linear Datalog to SQL for Program Analysis
本文提出一种将线性Datalog程序编译成递归SQL查询的方法，通过中间语言Midlog转换，利用现有数据库引擎执行程序分析，显著提高性能。
0 citationsRead paper
Selected work
Representative Papers
OTel: Building Domain-Specialized Telecom LLM Foundations for Intelligent Networks
This study addresses the suboptimal performance and resource scarcity of frontier AI in telecommunications by introducing OTel, the first unified open-source telecom AI resource. Through full-parameter post-training, instruction tuning, and safety alignment, we provide 30 baseline models spanning retrieval, reranking, and language modeling, alongside a comprehensive evaluation framework. Experimental results demonstrate state-of-the-art performance, achieving an NDCG@10 of 93.5% for embedding retrieval, an MRR@10 of 0.952 for reranking, and 88.2% accuracy for language models, with cumulative downloads exceeding 16 million. By filling a critical gap in the field, this work establishes a reproducible foundation of specialized large models for intelligent networks, effectively fostering community collaboration and accelerating technological advancement in telecom AI.
0 citationsRead paper
Time to Move on: Querying without Nulls and Bags
Traditional SQL undermines the theoretical foundations of the relational model by relying on nulls and bags, leading to semantic ambiguities and increased query complexity. This work proposes and implements Rel, a novel declarative query language that entirely eliminates nulls and bags, adhering strictly to set semantics and canonical relational algebra. Through an end-to-end system design and real-world deployment, we demonstrate that a null-free, bag-free relational system is not only expressively complete but also offers significant advantages in optimizability, semantic clarity, and engineering practicality. Our results confirm the feasibility and superiority of this paradigm for real-world applications.
0 citationsRead paper
A Compositional Language for Property Graphs
This work addresses the limited expressiveness of existing graph query languages—such as GQL and SQL/PGQ—which lack full compositionality and cannot capture complex path queries within the NLOGSPACE complexity class. To overcome this limitation, the paper introduces a novel query language that unifies graph pattern matching with relational querying through two key innovations: regular path queries enriched with variables and data-value comparisons, and a #Datalog-based graph transformation mechanism capable of constructing nodes, edges, and paths. This combination enables, for the first time, a systematically compositional approach to graph querying that precisely captures the full expressive power of NLOGSPACE. The proposed language not only resolves fundamental expressiveness gaps in current standards but also offers a practical and theoretically grounded extension pathway for both GQL and SQL/PGQ.
0 citationsRead paper
Neuro-Relational Programs: Unifying Queries and Neural Computation over Structured Data
This work addresses the challenge of unifying neural computation with logical querying in relational databases to enable structure-aware, learnable data processing. It proposes Neural Relational Programs (NRPs)—a declarative language that integrates Datalog-style rules with vector embedding operations, thereby embedding trainable neural components directly within a logical query framework for the first time. NRPs naturally subsume models such as graph neural networks (GNNs) and deep homomorphism networks: their nullary and unary fragments correspond respectively to non-adaptive queries and GNN-like architectures. Moreover, full NRPs precisely capture the expressivity of first-order logic with counting (FOCQ) over ordered databases and establish a theoretical connection to the circuit complexity class TC⁰.
0 citationsRead paper
Recent publications
Latest Papers
Compiling Linear Datalog to SQL for Program Analysis
本文提出一种将线性Datalog程序编译成递归SQL查询的方法，通过中间语言Midlog转换，利用现有数据库引擎执行程序分析，显著提高性能。
0 citationsRead paper
OTel: Building Domain-Specialized Telecom LLM Foundations for Intelligent Networks
This study addresses the suboptimal performance and resource scarcity of frontier AI in telecommunications by introducing OTel, the first unified open-source telecom AI resource. Through full-parameter post-training, instruction tuning, and safety alignment, we provide 30 baseline models spanning retrieval, reranking, and language modeling, alongside a comprehensive evaluation framework. Experimental results demonstrate state-of-the-art performance, achieving an NDCG@10 of 93.5% for embedding retrieval, an MRR@10 of 0.952 for reranking, and 88.2% accuracy for language models, with cumulative downloads exceeding 16 million. By filling a critical gap in the field, this work establishes a reproducible foundation of specialized large models for intelligent networks, effectively fostering community collaboration and accelerating technological advancement in telecom AI.
0 citationsRead paper
Time to Move on: Querying without Nulls and Bags
Traditional SQL undermines the theoretical foundations of the relational model by relying on nulls and bags, leading to semantic ambiguities and increased query complexity. This work proposes and implements Rel, a novel declarative query language that entirely eliminates nulls and bags, adhering strictly to set semantics and canonical relational algebra. Through an end-to-end system design and real-world deployment, we demonstrate that a null-free, bag-free relational system is not only expressively complete but also offers significant advantages in optimizability, semantic clarity, and engineering practicality. Our results confirm the feasibility and superiority of this paradigm for real-world applications.
0 citationsRead paper
A Compositional Language for Property Graphs
This work addresses the limited expressiveness of existing graph query languages—such as GQL and SQL/PGQ—which lack full compositionality and cannot capture complex path queries within the NLOGSPACE complexity class. To overcome this limitation, the paper introduces a novel query language that unifies graph pattern matching with relational querying through two key innovations: regular path queries enriched with variables and data-value comparisons, and a #Datalog-based graph transformation mechanism capable of constructing nodes, edges, and paths. This combination enables, for the first time, a systematically compositional approach to graph querying that precisely captures the full expressive power of NLOGSPACE. The proposed language not only resolves fundamental expressiveness gaps in current standards but also offers a practical and theoretically grounded extension pathway for both GQL and SQL/PGQ.
0 citationsRead paper
Neuro-Relational Programs: Unifying Queries and Neural Computation over Structured Data
This work addresses the challenge of unifying neural computation with logical querying in relational databases to enable structure-aware, learnable data processing. It proposes Neural Relational Programs (NRPs)—a declarative language that integrates Datalog-style rules with vector embedding operations, thereby embedding trainable neural components directly within a logical query framework for the first time. NRPs naturally subsume models such as graph neural networks (GNNs) and deep homomorphism networks: their nullary and unary fragments correspond respectively to non-adaptive queries and GNN-like architectures. Moreover, full NRPs precisely capture the expressivity of first-order logic with counting (FOCQ) over ordered databases and establish a theoretical connection to the circuit complexity class TC⁰.
0 citationsRead paper