Selected work

Representative Papers

MV2: Multi-View Multi-Vehicle Driving Dataset for Novel View Synthesis

Aug 12, 2026

Existing novel view synthesis methods are limited in real-world driving scenarios by sparse viewpoints, dynamic objects, and single-trajectory data. To address these challenges, this work introduces a multi-view, multi-vehicle urban driving dataset—captured synchronously from cars, scooters, and drones—that enables, for the first time, large-baseline image acquisition across vehicles and trajectories, accompanied by high-precision poses and pixel-level annotations. Sequences are registered via Structure-from-Motion (SfM) and refined with manually verified correspondences to support evaluation of differentiable rendering and novel view synthesis algorithms. Comprising 12,000 images across 50 scenes, the dataset’s benchmark experiments reveal the critical impact of viewpoint span on synthesis quality and highlight performance gaps in current pose estimation methods.

0 citationsRead paper

Towards trajectory-unsupervised physics-informed neural solvers for molecular dynamics

Aug 07, 2026

This work proposes DINaMo, a novel framework that introduces the first fully unsupervised neural trajectory solver for molecular dynamics. Unlike conventional neural approaches that rely on simulated trajectories, forces, or energies for supervised training and thus remain data-dependent, DINaMo models molecular trajectories as differentiable functions of time and trains exclusively through physical principles—namely Newton’s equations of motion, conservation laws, and analytically defined interaction potentials—without any reference simulation data. The method successfully reproduces key structural and dynamical observables in a Lennard-Jones argon system, including short-time coordinate evolution, energy conservation, and the radial distribution function of the liquid phase, thereby demonstrating the feasibility of learning molecular motion solely from physical constraints.

0 citationsRead paper

Active Learning Guided Design Space Refinement for Scalable Multi-Objective Bayesian Optimization in Materials Discovery

Aug 05, 2026

This work addresses the inefficiency of conventional Bayesian optimization in large-scale discrete material design spaces, where evaluation budgets are often wasted. To overcome this limitation, the authors propose an active learning–guided adaptive design space refinement mechanism integrated with multi-objective Bayesian optimization. This approach dynamically prunes the candidate space during iterative optimization while preserving Pareto front fidelity with high accuracy. Empirical results demonstrate substantial gains in efficiency: in case studies on COF-based methane/nitrogen separation and pressure vessel design, the method reduces the candidate space by approximately 50% while retaining over 99% of the original hypervolume. Furthermore, it accelerates early convergence and facilitates the discovery of high-quality Pareto-optimal solutions.

0 citationsRead paper

A Critical Analysis of Trustworthy AI Tools, Mark Frameworks, and the Implementation Chasms

Jul 16, 2026

This study addresses the significant gap between prevailing ethical principles for trustworthy artificial intelligence and their practical implementation, as existing frameworks often remain abstract and lack operational guidance. Leveraging the OECD dataset, this work presents the first systematic mapping and comparative analysis of global trustworthy AI tools and certification mechanisms, empirically examining dimensions such as ethical coverage, lifecycle integration, stakeholder engagement, and tool typology. The findings reveal an overemphasis on fairness, transparency, and robustness, while critical aspects like explainability, digital security, and environmental sustainability are largely neglected. Moreover, current tools predominantly target late-stage development phases, offering insufficient support for early design processes and educational policy. To bridge these gaps, the study proposes a governance pathway that broadens ethical objectives, spans the entire AI lifecycle, and strengthens multi-stakeholder collaboration, thereby offering structural insights for institutionalizing trustworthy AI.

0 citationsRead paper

AI Act Evaluation Benchmark: An Open, Transparent, and Reproducible Evaluation Dataset for NLP and RAG Systems

Mar 10, 2026

This work addresses the current lack of open resources for automated or semi-automated assessment of AI systems’ compliance with regulatory frameworks such as the EU AI Act, a gap that forces reliance on error-prone manual methods. To bridge this, the authors propose an open, transparent, and reproducible compliance evaluation dataset tailored for NLP and Retrieval-Augmented Generation (RAG) systems, encompassing four tasks: risk-level classification, provision retrieval, obligation generation, and question answering. Innovatively integrating legal domain knowledge with large language models, the approach enables grounded, high-relevance generation of controlled scenarios, effectively tackling the challenge of ambiguous risk boundaries—such as those between limited and minimal risk—specified in the Act. Experimental results demonstrate the dataset’s efficacy, achieving F1 scores of 0.87 and 0.85 on prohibited and high-risk scenarios, respectively, in RAG system evaluations.

0 citationsRead paper

Recent publications

Latest Papers

MV2: Multi-View Multi-Vehicle Driving Dataset for Novel View Synthesis

Aug 12, 2026

Existing novel view synthesis methods are limited in real-world driving scenarios by sparse viewpoints, dynamic objects, and single-trajectory data. To address these challenges, this work introduces a multi-view, multi-vehicle urban driving dataset—captured synchronously from cars, scooters, and drones—that enables, for the first time, large-baseline image acquisition across vehicles and trajectories, accompanied by high-precision poses and pixel-level annotations. Sequences are registered via Structure-from-Motion (SfM) and refined with manually verified correspondences to support evaluation of differentiable rendering and novel view synthesis algorithms. Comprising 12,000 images across 50 scenes, the dataset’s benchmark experiments reveal the critical impact of viewpoint span on synthesis quality and highlight performance gaps in current pose estimation methods.

0 citationsRead paper

Towards trajectory-unsupervised physics-informed neural solvers for molecular dynamics

Aug 07, 2026

This work proposes DINaMo, a novel framework that introduces the first fully unsupervised neural trajectory solver for molecular dynamics. Unlike conventional neural approaches that rely on simulated trajectories, forces, or energies for supervised training and thus remain data-dependent, DINaMo models molecular trajectories as differentiable functions of time and trains exclusively through physical principles—namely Newton’s equations of motion, conservation laws, and analytically defined interaction potentials—without any reference simulation data. The method successfully reproduces key structural and dynamical observables in a Lennard-Jones argon system, including short-time coordinate evolution, energy conservation, and the radial distribution function of the liquid phase, thereby demonstrating the feasibility of learning molecular motion solely from physical constraints.

0 citationsRead paper

Active Learning Guided Design Space Refinement for Scalable Multi-Objective Bayesian Optimization in Materials Discovery

Aug 05, 2026

This work addresses the inefficiency of conventional Bayesian optimization in large-scale discrete material design spaces, where evaluation budgets are often wasted. To overcome this limitation, the authors propose an active learning–guided adaptive design space refinement mechanism integrated with multi-objective Bayesian optimization. This approach dynamically prunes the candidate space during iterative optimization while preserving Pareto front fidelity with high accuracy. Empirical results demonstrate substantial gains in efficiency: in case studies on COF-based methane/nitrogen separation and pressure vessel design, the method reduces the candidate space by approximately 50% while retaining over 99% of the original hypervolume. Furthermore, it accelerates early convergence and facilitates the discovery of high-quality Pareto-optimal solutions.

0 citationsRead paper

A Critical Analysis of Trustworthy AI Tools, Mark Frameworks, and the Implementation Chasms

Jul 16, 2026

This study addresses the significant gap between prevailing ethical principles for trustworthy artificial intelligence and their practical implementation, as existing frameworks often remain abstract and lack operational guidance. Leveraging the OECD dataset, this work presents the first systematic mapping and comparative analysis of global trustworthy AI tools and certification mechanisms, empirically examining dimensions such as ethical coverage, lifecycle integration, stakeholder engagement, and tool typology. The findings reveal an overemphasis on fairness, transparency, and robustness, while critical aspects like explainability, digital security, and environmental sustainability are largely neglected. Moreover, current tools predominantly target late-stage development phases, offering insufficient support for early design processes and educational policy. To bridge these gaps, the study proposes a governance pathway that broadens ethical objectives, spans the entire AI lifecycle, and strengthens multi-stakeholder collaboration, thereby offering structural insights for institutionalizing trustworthy AI.

0 citationsRead paper

AI Act Evaluation Benchmark: An Open, Transparent, and Reproducible Evaluation Dataset for NLP and RAG Systems

Mar 10, 2026

This work addresses the current lack of open resources for automated or semi-automated assessment of AI systems’ compliance with regulatory frameworks such as the EU AI Act, a gap that forces reliance on error-prone manual methods. To bridge this, the authors propose an open, transparent, and reproducible compliance evaluation dataset tailored for NLP and Retrieval-Augmented Generation (RAG) systems, encompassing four tasks: risk-level classification, provision retrieval, obligation generation, and question answering. Innovatively integrating legal domain knowledge with large language models, the approach enables grounded, high-relevance generation of controlled scenarios, effectively tackling the challenge of ambiguous risk boundaries—such as those between limited and minimal risk—specified in the Act. Experimental results demonstrate the dataset’s efficacy, achieving F1 scores of 0.87 and 0.85 on prohibited and high-risk scenarios, respectively, in RAG system evaluations.

0 citationsRead paper