Selected work

Representative Papers

Target Trial Emulation with the R Package TTE: A Tutorial and Methodological Guide

Aug 02, 2026

This study addresses the susceptibility of causal effect estimation in observational studies to bias by proposing a systematic framework based on target trial emulation. By rigorously specifying eligibility criteria, treatment assignment, time zero, and follow-up rules to emulate a randomized controlled trial design, the approach integrates advanced statistical methods—including inverse probability weighting, weighted discrete-time survival models, model standardization, competing risk analysis, and cluster bootstrap at the individual level—into an end-to-end R implementation. The framework supports comparative analyses of both intention-to-treat and per-protocol effects and demonstrates its validity and practicality through two synthetic case studies, successfully estimating relative risks, absolute risks, and cumulative incidence functions.

0 citationsRead paper

Penalized likelihood inference for beta-binomial meta-analysis of proportions of rare events

Jul 28, 2026

This study addresses the finite-sample bias of maximum likelihood estimation (MLE) in meta-analyses of rare events, which arises due to data sparsity and studies reporting zero events. To mitigate this issue, the authors propose a maximum penalized likelihood estimation method within the beta-binomial random-effects model by incorporating a penalty term derived from Jeffreys’ prior. They further develop corresponding Wald-type and profile penalized likelihood confidence intervals. The proposed approach substantially enhances estimation stability and inferential reliability in sparse-data settings. Simulation results demonstrate that, particularly when the number of studies is small or event rates are low, the new method outperforms conventional MLE in terms of convergence, bias, and root mean squared error, while its confidence intervals maintain nominal coverage probabilities effectively.

0 citationsRead paper

Redesigning Regularization for Effective Policy Smoothing

Jun 11, 2026

This work addresses the practical limitations of existing regularization methods in reinforcement learning, which often fail to effectively smooth policy functions due to a disconnect between theory and implementation, leading to insufficient smoothness or compromised representational capacity. By identifying and rectifying three critical implementation flaws in prior approaches, the paper introduces a novel regularization mechanism that enforces both local and global Lipschitz continuity, thereby better aligning with theoretical expectations. The proposed method significantly enhances the smoothness of action outputs and control performance without sacrificing policy expressiveness. Its robustness to abrupt changes in target velocity is empirically validated across diverse tasks, algorithms, and sim-to-real experiments on quadrupedal robots.

0 citationsRead paper

Optimal Design Framework for Distributed Array Using Magnetically-Actuated Satellite Swarm

May 22, 2026

This study addresses the challenge of multi-constraint coupling in the design of distributed aperture antennas for electromagnetic formation flying. The authors propose a system-level design framework that unifies phased array performance requirements with constraints on satellite mass, power consumption, coil geometry, and formation-keeping dynamics into a single modeling framework. Notably, for the first time, formation-keeping metrics derived from distributed control simulations are incorporated into the aperture maximization problem, yielding a joint optimization model that accounts for aperture size, power allocation, coil parameters, and sidelobe envelope specifications. Leveraging a static mesh reference structure, the framework efficiently computes feasible apertures under fixed system mass. Case studies demonstrate that at a 0.15 m inter-satellite spacing, power generation and coil geometry dominate the design constraints, whereas at 0.60 m, coil loading tends to exceed limits—validating the framework’s capability to effectively evaluate and optimize aperture configurations under complex, coupled constraints.

0 citationsRead paper

Refining and Reusing Annotation Guidelines for LLM Annotation

May 20, 2026

This work addresses the challenge that large language models (LLMs) struggle to adhere to domain-specific gold-standard annotation guidelines in zero-shot settings. To mitigate this limitation, the authors propose a mediation framework that iteratively reuses and refines annotation guidelines, introducing guideline evolution as a novel alignment mechanism to enhance annotation consistency and accuracy under low-supervision conditions. The approach integrates reasoning-optimized variants from three major LLM families—GPT, Gemini, and DeepSeek—and employs iterative guideline consolidation and fine-tuning. Evaluated on biomedical named entity recognition benchmarks including NCBI Disease, BC5CDR, and BioRED, the method demonstrates significant improvements in the models’ compliance with expert annotation standards.

0 citationsRead paper

Recent publications

Latest Papers

Target Trial Emulation with the R Package TTE: A Tutorial and Methodological Guide

Aug 02, 2026

This study addresses the susceptibility of causal effect estimation in observational studies to bias by proposing a systematic framework based on target trial emulation. By rigorously specifying eligibility criteria, treatment assignment, time zero, and follow-up rules to emulate a randomized controlled trial design, the approach integrates advanced statistical methods—including inverse probability weighting, weighted discrete-time survival models, model standardization, competing risk analysis, and cluster bootstrap at the individual level—into an end-to-end R implementation. The framework supports comparative analyses of both intention-to-treat and per-protocol effects and demonstrates its validity and practicality through two synthetic case studies, successfully estimating relative risks, absolute risks, and cumulative incidence functions.

0 citationsRead paper

Penalized likelihood inference for beta-binomial meta-analysis of proportions of rare events

Jul 28, 2026

This study addresses the finite-sample bias of maximum likelihood estimation (MLE) in meta-analyses of rare events, which arises due to data sparsity and studies reporting zero events. To mitigate this issue, the authors propose a maximum penalized likelihood estimation method within the beta-binomial random-effects model by incorporating a penalty term derived from Jeffreys’ prior. They further develop corresponding Wald-type and profile penalized likelihood confidence intervals. The proposed approach substantially enhances estimation stability and inferential reliability in sparse-data settings. Simulation results demonstrate that, particularly when the number of studies is small or event rates are low, the new method outperforms conventional MLE in terms of convergence, bias, and root mean squared error, while its confidence intervals maintain nominal coverage probabilities effectively.

0 citationsRead paper

Redesigning Regularization for Effective Policy Smoothing

Jun 11, 2026

This work addresses the practical limitations of existing regularization methods in reinforcement learning, which often fail to effectively smooth policy functions due to a disconnect between theory and implementation, leading to insufficient smoothness or compromised representational capacity. By identifying and rectifying three critical implementation flaws in prior approaches, the paper introduces a novel regularization mechanism that enforces both local and global Lipschitz continuity, thereby better aligning with theoretical expectations. The proposed method significantly enhances the smoothness of action outputs and control performance without sacrificing policy expressiveness. Its robustness to abrupt changes in target velocity is empirically validated across diverse tasks, algorithms, and sim-to-real experiments on quadrupedal robots.

0 citationsRead paper

Optimal Design Framework for Distributed Array Using Magnetically-Actuated Satellite Swarm

May 22, 2026

This study addresses the challenge of multi-constraint coupling in the design of distributed aperture antennas for electromagnetic formation flying. The authors propose a system-level design framework that unifies phased array performance requirements with constraints on satellite mass, power consumption, coil geometry, and formation-keeping dynamics into a single modeling framework. Notably, for the first time, formation-keeping metrics derived from distributed control simulations are incorporated into the aperture maximization problem, yielding a joint optimization model that accounts for aperture size, power allocation, coil parameters, and sidelobe envelope specifications. Leveraging a static mesh reference structure, the framework efficiently computes feasible apertures under fixed system mass. Case studies demonstrate that at a 0.15 m inter-satellite spacing, power generation and coil geometry dominate the design constraints, whereas at 0.60 m, coil loading tends to exceed limits—validating the framework’s capability to effectively evaluate and optimize aperture configurations under complex, coupled constraints.

0 citationsRead paper

Refining and Reusing Annotation Guidelines for LLM Annotation

May 20, 2026

This work addresses the challenge that large language models (LLMs) struggle to adhere to domain-specific gold-standard annotation guidelines in zero-shot settings. To mitigate this limitation, the authors propose a mediation framework that iteratively reuses and refines annotation guidelines, introducing guideline evolution as a novel alignment mechanism to enhance annotation consistency and accuracy under low-supervision conditions. The approach integrates reasoning-optimized variants from three major LLM families—GPT, Gemini, and DeepSeek—and employs iterative guideline consolidation and fine-tuning. Evaluated on biomedical named entity recognition benchmarks including NCBI Disease, BC5CDR, and BioRED, the method demonstrates significant improvements in the models’ compliance with expert annotation standards.

0 citationsRead paper