Selected work

Representative Papers

A Framework for Managing the Models of Engineered Quantum Systems

Jul 11, 2026

Existing model synchronization approaches struggle to accommodate the superposition and entanglement inherent in quantum systems, thereby failing to ensure consistency across heterogeneous models from multiple domains. To address this challenge, this work proposes QSysMM—the first dedicated model management framework tailored for engineered quantum systems—which establishes a unified digital single source of truth by harmonizing quantum engineering models along four dimensions: ontology, abstraction, composition, and exposure. Built upon the SysML v2 technology stack, we introduce QSysML, a companion modeling language that integrates model-driven engineering with quantum information science to preserve quantum semantics throughout model transformations. This framework provides a comprehensive foundation for the construction, verification, and maintenance of complex quantum systems through robust model synchronization and management.

0 citationsRead paper

Enhanced Neural Video Representation Compression across Extreme Complexity and Quality Scales

Jun 26, 2026

Existing neural video codecs struggle to balance complexity and scalability: lightweight models suffer from limited rate-distortion performance, while high-performance models exhibit rapidly increasing computational complexity with quality, and lack a unified architecture for multi-scenario deployment. This work proposes NVRC++, a unified neural video coding framework based on implicit neural representations (INRs). By integrating a lightweight INR, multi-resolution feature grids, an efficient overfitting-aware optimization tailored for long videos, and an advanced high-dimensional entropy model, NVRC++ supports four complexity levels (7k–360k MACs/pixel) within a single fixed architecture. Each level spans a broad bitrate range while maintaining real-time decoding. Compared to the state-of-the-art NVRC, NVRC++ achieves up to 7.6× faster decoding at comparable rate-distortion performance, marking the first neural video compression system capable of efficient operation across extreme scales of both complexity and quality.

0 citationsRead paper

A Low-Latency Semantic State Estimator using Latent Predictive Learning for Dynamic Network Monitoring and Orchestration

Jun 07, 2026

In dynamic cloud-edge environments, conventional network monitoring approaches struggle to simultaneously achieve low latency, scalability, and semantic understanding. This work proposes a Latent-variable Predictive State Estimator (LPSE) that leverages topology-adaptive temporal encoding to map telemetry data from variable-scale nodes into permutation-invariant slot-based representations, enabling fixed-overhead, single-pass inference through a semantic codebook. LPSE is the first method capable of generalizing to node additions, removals, and reordering without requiring retraining, substantially enhancing dynamic adaptability. Experimental results on multi-node Kubernetes clusters demonstrate that LPSE attains a semantic prediction accuracy of 82.42%, while reducing average inference latency by 41× and memory footprint by 15× compared to deployable 4B-parameter large language model endpoints.

0 citationsRead paper

Graph Transformers and Stabilized Reinforcement Learning for Large-Scale Dynamic Routing Modulation and Spectrum Allocation in Elastic Optical Networks

May 03, 2026

This work addresses the lack of scalable and stable reinforcement learning solutions for the routing, modulation, and spectrum assignment (RMSA) problem in large-scale dynamic optical networks. It proposes the first application of the Transformer architecture to dynamic RMSA, integrating graph-structured rotary position encoding, off-policy invalid-action masking, and effective-quality regularization to establish a robust and efficient reinforcement learning training framework. Evaluated on a real-world network topology with 143 nodes and 362 links, the method supports up to 13% additional traffic load and achieves a 4% improvement over the best-performing baseline while maintaining a blocking probability below 0.1%. The authors release all code publicly to facilitate reproducibility and future research.

0 citationsRead paper

Benchmarking and Evaluating VLMs for Software Architecture Diagram Understanding

Apr 05, 2026

This work addresses the limited capability of existing vision-language models (VLMs) in comprehending software architecture diagrams—structured engineering artifacts—and the absence of dedicated evaluation benchmarks. To bridge this gap, we introduce SADU, the first VLM benchmark tailored to the software design phase, comprising 154 architecture diagrams of behavioral, structural, and entity-relationship types along with 2,431 multimodal question-answering tasks. We systematically evaluate 11 prominent VLMs, including Gemini, GPT, Claude, and Qwen, on tasks such as counting and retrieval-based reasoning. The top-performing model, gemini-3-flash-preview, achieves only 70.18% accuracy, revealing significant limitations in current VLMs’ ability to parse diagrammatic structures and localize visual relationships. This benchmark establishes a new paradigm for advancing VLM research in software engineering contexts.

0 citationsRead paper

Recent publications

Latest Papers

A Framework for Managing the Models of Engineered Quantum Systems

Jul 11, 2026

Existing model synchronization approaches struggle to accommodate the superposition and entanglement inherent in quantum systems, thereby failing to ensure consistency across heterogeneous models from multiple domains. To address this challenge, this work proposes QSysMM—the first dedicated model management framework tailored for engineered quantum systems—which establishes a unified digital single source of truth by harmonizing quantum engineering models along four dimensions: ontology, abstraction, composition, and exposure. Built upon the SysML v2 technology stack, we introduce QSysML, a companion modeling language that integrates model-driven engineering with quantum information science to preserve quantum semantics throughout model transformations. This framework provides a comprehensive foundation for the construction, verification, and maintenance of complex quantum systems through robust model synchronization and management.

0 citationsRead paper

Enhanced Neural Video Representation Compression across Extreme Complexity and Quality Scales

Jun 26, 2026

Existing neural video codecs struggle to balance complexity and scalability: lightweight models suffer from limited rate-distortion performance, while high-performance models exhibit rapidly increasing computational complexity with quality, and lack a unified architecture for multi-scenario deployment. This work proposes NVRC++, a unified neural video coding framework based on implicit neural representations (INRs). By integrating a lightweight INR, multi-resolution feature grids, an efficient overfitting-aware optimization tailored for long videos, and an advanced high-dimensional entropy model, NVRC++ supports four complexity levels (7k–360k MACs/pixel) within a single fixed architecture. Each level spans a broad bitrate range while maintaining real-time decoding. Compared to the state-of-the-art NVRC, NVRC++ achieves up to 7.6× faster decoding at comparable rate-distortion performance, marking the first neural video compression system capable of efficient operation across extreme scales of both complexity and quality.

0 citationsRead paper

A Low-Latency Semantic State Estimator using Latent Predictive Learning for Dynamic Network Monitoring and Orchestration

Jun 07, 2026

In dynamic cloud-edge environments, conventional network monitoring approaches struggle to simultaneously achieve low latency, scalability, and semantic understanding. This work proposes a Latent-variable Predictive State Estimator (LPSE) that leverages topology-adaptive temporal encoding to map telemetry data from variable-scale nodes into permutation-invariant slot-based representations, enabling fixed-overhead, single-pass inference through a semantic codebook. LPSE is the first method capable of generalizing to node additions, removals, and reordering without requiring retraining, substantially enhancing dynamic adaptability. Experimental results on multi-node Kubernetes clusters demonstrate that LPSE attains a semantic prediction accuracy of 82.42%, while reducing average inference latency by 41× and memory footprint by 15× compared to deployable 4B-parameter large language model endpoints.

0 citationsRead paper

Graph Transformers and Stabilized Reinforcement Learning for Large-Scale Dynamic Routing Modulation and Spectrum Allocation in Elastic Optical Networks

May 03, 2026

This work addresses the lack of scalable and stable reinforcement learning solutions for the routing, modulation, and spectrum assignment (RMSA) problem in large-scale dynamic optical networks. It proposes the first application of the Transformer architecture to dynamic RMSA, integrating graph-structured rotary position encoding, off-policy invalid-action masking, and effective-quality regularization to establish a robust and efficient reinforcement learning training framework. Evaluated on a real-world network topology with 143 nodes and 362 links, the method supports up to 13% additional traffic load and achieves a 4% improvement over the best-performing baseline while maintaining a blocking probability below 0.1%. The authors release all code publicly to facilitate reproducibility and future research.

0 citationsRead paper

Benchmarking and Evaluating VLMs for Software Architecture Diagram Understanding

Apr 05, 2026

This work addresses the limited capability of existing vision-language models (VLMs) in comprehending software architecture diagrams—structured engineering artifacts—and the absence of dedicated evaluation benchmarks. To bridge this gap, we introduce SADU, the first VLM benchmark tailored to the software design phase, comprising 154 architecture diagrams of behavioral, structural, and entity-relationship types along with 2,431 multimodal question-answering tasks. We systematically evaluate 11 prominent VLMs, including Gemini, GPT, Claude, and Qwen, on tasks such as counting and retrieval-based reasoning. The top-performing model, gemini-3-flash-preview, achieves only 70.18% accuracy, revealing significant limitations in current VLMs’ ability to parse diagrammatic structures and localize visual relationships. This benchmark establishes a new paradigm for advancing VLM research in software engineering contexts.

0 citationsRead paper