Selected work

Representative Papers

Graph-Based Agentic AI with LangGraph: Workflow Pathways for Long-Running Stateful Business Processes

Jul 21, 2026

This work addresses the challenges of constructing and managing generative AI agent systems for long-horizon, stateful, multi-step business processes by proposing a graph-structured workflow design methodology. Leveraging the LangGraph framework, it explicitly models core mechanisms such as state management, conditional routing, and human-in-the-loop interventions. The approach is instantiated in three representative applications: SQL analysis with repair loops, retrieval-augmented generation gated by evidential validation, and human-AI collaborative policy review supporting interruption and checkpoint-based recovery. By treating behaviors like routing, pausing, and audit trails as explicit product features rather than implicit prompt logic, this study not only delineates the applicability boundaries of LangGraph in high-complexity workflows but also substantially enhances system controllability, reliability, and auditability in real-world operational settings, establishing a reusable engineering paradigm.

0 citationsRead paper

Structural Divergence of the Roman--Byzantine Trade Network, 0--1453\,CE: Persistent Homology, Topological Velocity, and Criticality Indicators of Imperial Collapse

Jul 06, 2026

This study uncovers the topological linkage between the structural evolution of the Roman–Byzantine trade network over a millennium and imperial collapse, overcoming misinterpretations in historical network analysis caused by data sparsity and hub bias. Leveraging the ORBIS geospatial network—comprising 2,599 nodes and 4,503 trimodal edges—the work introduces, for the first time, a universal indicator of imperial collapse termed the “topological percolation threshold,” derived from persistent homology, Wasserstein distance, and topological velocity. Through hierarchical network decoupling and entropy analysis, it reveals a steadily widening entropy gap between eastern and western subnetworks since 1 CE, identifies 495 CE as the peak of topological velocity and 620 CE as the height of economic–geographic decoupling, and confirms that both the collapses of 476 CE and 1453 CE correspond to a critical order parameter H* ≈ 0.524.

0 citationsRead paper

Adressing Separation: A Firth-corrected Joint Model for Longitudinal and Time-to-event Data with an Application on Dropout from Vocational Training

Jun 09, 2026

This study addresses the bias in parameter estimation arising from complete separation when modeling categorical covariates within joint models for longitudinal and time-to-event data. To mitigate this issue, the authors introduce Firth’s penalized likelihood into the joint modeling framework for the first time, embedding it within an Expectation–Maximization (EM) algorithm. The proposed approach effectively alleviates estimation failure under separation and substantially reduces parameter bias. Extensive simulations and an analysis of real-world data from a German vocational training program demonstrate that the method yields robust estimates of covariate effects and uncovers both direct and indirect influences of socioeconomic factors on dropout risk. This work thus provides a reliable frequentist solution for joint modeling scenarios afflicted by complete separation.

0 citationsRead paper

LLM-supported document separation for printed reviews from zbMATH Open

Apr 01, 2026

This study addresses the challenge of rendering approximately 831,000 scanned mathematical documents in zbMATH Open machine-readable, a task hindered by mixed-document layouts, cross-page content, and complex formatting. The authors propose a document segmentation pipeline that integrates OCR with fine-tuned large language models (LLMs). The approach first employs Mathpix OCR to achieve high-quality LaTeX conversion, followed by an ensemble of multiple fine-tuned LLMs whose predictions are combined via majority voting to accurately identify document boundaries. This work represents the first application of fine-tuned LLMs coupled with majority voting for mathematical document segmentation and includes a systematic evaluation of various OCR tools on LaTeX reconstruction performance. Experimental results demonstrate an overall accuracy of 89.1% in boundary detection, successfully processing 810,977 documents—of which 721,288 achieved precise LaTeX-formatted boundaries—significantly outperforming rule-based, general-purpose LLM, and computer vision baselines.

0 citationsRead paper

Macroscopic Signatures of Gauge-Mediated Contagion: Deriving Behavioral Shielding from Stochastic Field Theory

Mar 31, 2026

This work proposes a unified framework for modeling the coupling between epidemic dynamics and spontaneous behavioral responses by starting from microscopic stochastic transmission mechanisms. Building on the Doi–Peliti stochastic field theory, pathogens are represented as gauge-mediated fields, while a reactive immunity field capable of spontaneous symmetry breaking is introduced. Macroscopic reaction–diffusion equations are derived via a saddle-point approximation of the effective action. The study innovatively maps concepts from quantum field theory—such as the Coleman–Weinberg mechanism, Debye screening, and vacuum polarization—onto epidemiological contexts, revealing how fear-induced drift and cubic screening dynamically suppress the effective reproduction number. The model is validated against high-resolution COVID-19 data from Germany, accurately reproducing both the susceptibility-density-squared dependence of the free energy and the behaviorally induced cubic nonlinear suppression term.

0 citationsRead paper

Recent publications

Latest Papers

Graph-Based Agentic AI with LangGraph: Workflow Pathways for Long-Running Stateful Business Processes

Jul 21, 2026

This work addresses the challenges of constructing and managing generative AI agent systems for long-horizon, stateful, multi-step business processes by proposing a graph-structured workflow design methodology. Leveraging the LangGraph framework, it explicitly models core mechanisms such as state management, conditional routing, and human-in-the-loop interventions. The approach is instantiated in three representative applications: SQL analysis with repair loops, retrieval-augmented generation gated by evidential validation, and human-AI collaborative policy review supporting interruption and checkpoint-based recovery. By treating behaviors like routing, pausing, and audit trails as explicit product features rather than implicit prompt logic, this study not only delineates the applicability boundaries of LangGraph in high-complexity workflows but also substantially enhances system controllability, reliability, and auditability in real-world operational settings, establishing a reusable engineering paradigm.

0 citationsRead paper

Structural Divergence of the Roman--Byzantine Trade Network, 0--1453\,CE: Persistent Homology, Topological Velocity, and Criticality Indicators of Imperial Collapse

Jul 06, 2026

This study uncovers the topological linkage between the structural evolution of the Roman–Byzantine trade network over a millennium and imperial collapse, overcoming misinterpretations in historical network analysis caused by data sparsity and hub bias. Leveraging the ORBIS geospatial network—comprising 2,599 nodes and 4,503 trimodal edges—the work introduces, for the first time, a universal indicator of imperial collapse termed the “topological percolation threshold,” derived from persistent homology, Wasserstein distance, and topological velocity. Through hierarchical network decoupling and entropy analysis, it reveals a steadily widening entropy gap between eastern and western subnetworks since 1 CE, identifies 495 CE as the peak of topological velocity and 620 CE as the height of economic–geographic decoupling, and confirms that both the collapses of 476 CE and 1453 CE correspond to a critical order parameter H* ≈ 0.524.

0 citationsRead paper

Adressing Separation: A Firth-corrected Joint Model for Longitudinal and Time-to-event Data with an Application on Dropout from Vocational Training

Jun 09, 2026

This study addresses the bias in parameter estimation arising from complete separation when modeling categorical covariates within joint models for longitudinal and time-to-event data. To mitigate this issue, the authors introduce Firth’s penalized likelihood into the joint modeling framework for the first time, embedding it within an Expectation–Maximization (EM) algorithm. The proposed approach effectively alleviates estimation failure under separation and substantially reduces parameter bias. Extensive simulations and an analysis of real-world data from a German vocational training program demonstrate that the method yields robust estimates of covariate effects and uncovers both direct and indirect influences of socioeconomic factors on dropout risk. This work thus provides a reliable frequentist solution for joint modeling scenarios afflicted by complete separation.

0 citationsRead paper

LLM-supported document separation for printed reviews from zbMATH Open

Apr 01, 2026

This study addresses the challenge of rendering approximately 831,000 scanned mathematical documents in zbMATH Open machine-readable, a task hindered by mixed-document layouts, cross-page content, and complex formatting. The authors propose a document segmentation pipeline that integrates OCR with fine-tuned large language models (LLMs). The approach first employs Mathpix OCR to achieve high-quality LaTeX conversion, followed by an ensemble of multiple fine-tuned LLMs whose predictions are combined via majority voting to accurately identify document boundaries. This work represents the first application of fine-tuned LLMs coupled with majority voting for mathematical document segmentation and includes a systematic evaluation of various OCR tools on LaTeX reconstruction performance. Experimental results demonstrate an overall accuracy of 89.1% in boundary detection, successfully processing 810,977 documents—of which 721,288 achieved precise LaTeX-formatted boundaries—significantly outperforming rule-based, general-purpose LLM, and computer vision baselines.

0 citationsRead paper

Macroscopic Signatures of Gauge-Mediated Contagion: Deriving Behavioral Shielding from Stochastic Field Theory

Mar 31, 2026

This work proposes a unified framework for modeling the coupling between epidemic dynamics and spontaneous behavioral responses by starting from microscopic stochastic transmission mechanisms. Building on the Doi–Peliti stochastic field theory, pathogens are represented as gauge-mediated fields, while a reactive immunity field capable of spontaneous symmetry breaking is introduced. Macroscopic reaction–diffusion equations are derived via a saddle-point approximation of the effective action. The study innovatively maps concepts from quantum field theory—such as the Coleman–Weinberg mechanism, Debye screening, and vacuum polarization—onto epidemiological contexts, revealing how fear-induced drift and cubic screening dynamically suppress the effective reproduction number. The model is validated against high-resolution COVID-19 data from Germany, accurately reproducing both the susceptibility-density-squared dependence of the free energy and the behaviorally induced cubic nonlinear suppression term.

0 citationsRead paper