FIRM-WM: State-factorized factual-interventional recurrent modeling for reward-free visual planning
研究提出FIRM-WM模型，通过分离目标可比配置与动态预测纤维解决无奖励视觉规划中的状态表示问题，并在多个任务上实现更高效能。
0 citationsRead paper
Selected work
Representative Papers
Communication-efficient parallel Bruhat decomposition
This work addresses the high communication and synchronization overhead inherent in Bruhat decomposition under the Bulk-Synchronous Parallel (BSP) model by proposing a novel block-recursive parallel algorithm. For the first time, this approach achieves communication-efficient Bruhat decomposition within the BSP framework, significantly reducing both communication complexity and the number of synchronization steps while preserving numerical stability and matrix structural symmetry. Experimental results demonstrate that the proposed algorithm substantially enhances the scalability of Bruhat decomposition on general-purpose parallel platforms, thereby providing an efficient foundation for large-scale sparse linear algebra computations.
0 citationsRead paper
Fast Discovery of Inclusion Dependencies with Desbordante
This work addresses the high computational cost and insufficient engineering optimization in inclusion dependency discovery by efficiently implementing the Spider and Faida algorithms within the Desbordante system. Through implementation-level optimizations—including parallelization, SIMD vectorization, custom hash tables, and data caching—it systematically integrates algorithmic design with engineering practice for the first time. Specifically, a memory-efficient, highly parallel variant is introduced for the classic Spider algorithm, while four performance-enhancing optimizations are incorporated into the state-of-the-art approximate algorithm Faida. Compared to the Java-based Metanome framework, the optimized implementations achieve up to a 5× speedup for Spider and an 8× speedup for Faida, accompanied by substantially reduced memory consumption.
0 citationsRead paper
FastGFDs: Efficient Validation of Graph Functional Dependencies with Desbordante
This work addresses the high computational cost of graph functional dependency (GFD) validation, which typically relies on high-performance clusters and is inefficient on commodity hardware. To overcome this limitation, we propose FastGFDs, the first algorithm to integrate Core-First decomposition and Compact Path Indexing (CPI) into GFD verification. FastGFDs introduces an efficient sequential execution strategy tailored for single-node, low-resource environments, combining full-graph processing with an optimized subgraph matching mechanism. Implemented within the Desbordante framework, our approach achieves an average speedup of 2.6× (up to 3×) on real-world graph datasets while reducing memory consumption to one-fifth of existing methods. We also release the first open-source tool for GFD validation, making this capability accessible to a broader research community.
0 citationsRead paper
Order in Desbordante: Techniques for Efficient Implementation of Order Dependency Discovery Algorithms
This work addresses the critical role of order dependencies (ODs) in query optimization and data cleaning, where existing discovery algorithms suffer from efficiency and scalability limitations. The study presents the first systematic analysis of engineering bottlenecks in OD discovery and introduces a high-performance implementation by re-engineering the FASTOD and ORDER algorithms in C++. These optimized algorithms are integrated into the Desbordante data profiling framework, leveraging advanced memory management and computational strategies. Compared to the original implementations, the proposed approach achieves up to a 10× speedup and reduces memory consumption by as much as 2.9×, substantially enhancing the practicality and scalability of OD discovery for real-world applications.
0 citationsRead paper
Recent publications
Latest Papers
FIRM-WM: State-factorized factual-interventional recurrent modeling for reward-free visual planning
研究提出FIRM-WM模型，通过分离目标可比配置与动态预测纤维解决无奖励视觉规划中的状态表示问题，并在多个任务上实现更高效能。
0 citationsRead paper
Communication-efficient parallel Bruhat decomposition
This work addresses the high communication and synchronization overhead inherent in Bruhat decomposition under the Bulk-Synchronous Parallel (BSP) model by proposing a novel block-recursive parallel algorithm. For the first time, this approach achieves communication-efficient Bruhat decomposition within the BSP framework, significantly reducing both communication complexity and the number of synchronization steps while preserving numerical stability and matrix structural symmetry. Experimental results demonstrate that the proposed algorithm substantially enhances the scalability of Bruhat decomposition on general-purpose parallel platforms, thereby providing an efficient foundation for large-scale sparse linear algebra computations.
0 citationsRead paper
Fast Discovery of Inclusion Dependencies with Desbordante
This work addresses the high computational cost and insufficient engineering optimization in inclusion dependency discovery by efficiently implementing the Spider and Faida algorithms within the Desbordante system. Through implementation-level optimizations—including parallelization, SIMD vectorization, custom hash tables, and data caching—it systematically integrates algorithmic design with engineering practice for the first time. Specifically, a memory-efficient, highly parallel variant is introduced for the classic Spider algorithm, while four performance-enhancing optimizations are incorporated into the state-of-the-art approximate algorithm Faida. Compared to the Java-based Metanome framework, the optimized implementations achieve up to a 5× speedup for Spider and an 8× speedup for Faida, accompanied by substantially reduced memory consumption.
0 citationsRead paper
FastGFDs: Efficient Validation of Graph Functional Dependencies with Desbordante
This work addresses the high computational cost of graph functional dependency (GFD) validation, which typically relies on high-performance clusters and is inefficient on commodity hardware. To overcome this limitation, we propose FastGFDs, the first algorithm to integrate Core-First decomposition and Compact Path Indexing (CPI) into GFD verification. FastGFDs introduces an efficient sequential execution strategy tailored for single-node, low-resource environments, combining full-graph processing with an optimized subgraph matching mechanism. Implemented within the Desbordante framework, our approach achieves an average speedup of 2.6× (up to 3×) on real-world graph datasets while reducing memory consumption to one-fifth of existing methods. We also release the first open-source tool for GFD validation, making this capability accessible to a broader research community.
0 citationsRead paper
Order in Desbordante: Techniques for Efficient Implementation of Order Dependency Discovery Algorithms
This work addresses the critical role of order dependencies (ODs) in query optimization and data cleaning, where existing discovery algorithms suffer from efficiency and scalability limitations. The study presents the first systematic analysis of engineering bottlenecks in OD discovery and introduces a high-performance implementation by re-engineering the FASTOD and ORDER algorithms in C++. These optimized algorithms are integrated into the Desbordante data profiling framework, leveraging advanced memory management and computational strategies. Compared to the original implementations, the proposed approach achieves up to a 10× speedup and reduces memory consumption by as much as 2.9×, substantially enhancing the practicality and scalability of OD discovery for real-world applications.
0 citationsRead paper