Selected work

Representative Papers

World-Model-Aware Responsibility Allocation in Heterogeneous Logistics Systems

Jul 16, 2026

This study addresses decision conflicts and deadlocks arising from inconsistent world models between a central controller and autonomous devices in mixed-autonomy logistics systems, a challenge poorly handled by conventional scheduling approaches. The authors propose the World-Model-Aware Responsibility Framework (WMARF), which uniquely leverages the quality of world models as a basis for dynamic authority delegation. By continuously adjusting decision-making authority according to device autonomy levels and classifying authority states to proactively identify and avoid deadlocks, WMARF overcomes the limitations of static permission schemes and supports evolving system autonomy. Validated through discrete-event simulation adhering to the VDA 5050 standard and employing a proximity-triggered authority handover mechanism, the framework successfully prevents model-discrepancy-induced deadlocks in scenarios where two autonomous vehicles approach a semi-automated transfer point under static control.

0 citationsRead paper

On-the-Fly Fine-Tuning of Foundational Neural Network Potentials: A Bayesian Neural Network Approach

Jul 18, 2025

Constructing high-accuracy force fields for molecular dynamics is hindered by prohibitive data acquisition costs, poor modeling of rare events (e.g., transition states), and the absence of uncertainty quantification during fine-tuning. Method: This work proposes an online fine-tuning framework based on Bayesian neural networks (BNNs), the first to integrate BNNs into the dynamic updating of foundational neural network potentials. The method estimates predictive uncertainty in real time during fine-tuning and autonomously triggers first-principles calculations to augment critical configurations—such as transition-state geometries—enabling closed-loop active learning. Contribution/Results: Compared with conventional fine-tuning, our approach drastically reduces reliance on labeled data and cuts first-principles computational cost by several-fold while preserving accuracy. It demonstrates superior generalization and sampling efficiency in simulating complex systems and rare-event processes.

0 citationsRead paper

Recent publications

Latest Papers

World-Model-Aware Responsibility Allocation in Heterogeneous Logistics Systems

Jul 16, 2026

This study addresses decision conflicts and deadlocks arising from inconsistent world models between a central controller and autonomous devices in mixed-autonomy logistics systems, a challenge poorly handled by conventional scheduling approaches. The authors propose the World-Model-Aware Responsibility Framework (WMARF), which uniquely leverages the quality of world models as a basis for dynamic authority delegation. By continuously adjusting decision-making authority according to device autonomy levels and classifying authority states to proactively identify and avoid deadlocks, WMARF overcomes the limitations of static permission schemes and supports evolving system autonomy. Validated through discrete-event simulation adhering to the VDA 5050 standard and employing a proximity-triggered authority handover mechanism, the framework successfully prevents model-discrepancy-induced deadlocks in scenarios where two autonomous vehicles approach a semi-automated transfer point under static control.

0 citationsRead paper

On-the-Fly Fine-Tuning of Foundational Neural Network Potentials: A Bayesian Neural Network Approach

Jul 18, 2025

Constructing high-accuracy force fields for molecular dynamics is hindered by prohibitive data acquisition costs, poor modeling of rare events (e.g., transition states), and the absence of uncertainty quantification during fine-tuning. Method: This work proposes an online fine-tuning framework based on Bayesian neural networks (BNNs), the first to integrate BNNs into the dynamic updating of foundational neural network potentials. The method estimates predictive uncertainty in real time during fine-tuning and autonomously triggers first-principles calculations to augment critical configurations—such as transition-state geometries—enabling closed-loop active learning. Contribution/Results: Compared with conventional fine-tuning, our approach drastically reduces reliance on labeled data and cuts first-principles computational cost by several-fold while preserving accuracy. It demonstrates superior generalization and sampling efficiency in simulating complex systems and rare-event processes.

0 citationsRead paper