Debiasing as a Measurement Intervention: Calibrated Ties and Resolution Loss in LLM-as-a-Judge Evaluation
研究通过TraceJudgeBench基准测试评估LLM作为评判时去除引用偏见的方法，发现强去偏指令虽然能减少偏见但会损害分辨能力，并提出了一种分离技术以恢复分辨率。
0 citationsRead paper
Selected work
Representative Papers
Can Large Language Models Explain Flight Safety Events? A Prior-Guided Semantic LLM-based Approach
研究提出FlightLLM，一种基于先验引导的语义大语言模型方法，通过特征工程、语义离散化及少量样本对比学习等手段解决飞行安全事件解释问题。
0 citationsRead paper
StatCite: A Large-scale Citation Network Dataset for Statistics and Data Science
This study addresses the notable scarcity of large-scale, structurally deep, and metadata-rich citation network datasets in statistics and data science. To bridge this gap, the authors construct a comprehensive citation network dataset comprising 189,101 papers published between 1981 and 2025, enriched with extensive metadata including titles, authors, abstracts, keywords, and references. For the first time in this domain, the work simultaneously integrates four complementary network types: paper citation, co-citation, bibliographic coupling, and journal citation networks. Through systematic data collection, network construction, and community detection methodologies, the dataset successfully reproduces canonical structural properties of citation networks and reveals multiple core research themes. This resource provides a high-quality foundation for advancing bibliometric analysis, knowledge graph development, and science of science research.
0 citationsRead paper
Quantitative Stability and Contraction Principles for Mean-Field G-SDEs
This study establishes a quantitative stability theory for mean-field stochastic differential equations driven by G-Brownian motion (mean-field G-SDEs) under non-Lipschitz coefficients, volatility uncertainty, and square-integrable initial data. Methodologically, we develop a novel Bihari–Osgood-type inequality within the G-expectation framework, integrating nonlinear expectation theory with mean-field analysis to derive explicit stability moduli with respect to initial conditions and coefficient perturbations; we further identify and formalize a short-time contraction property. Key contributions include: (1) the first precise quantification of solution-map sensitivity under coupled non-Lipschitz and volatility-uncertain dynamics; (2) a global stability propagation mechanism; and (3) sharp Hölder continuity estimates for the data-to-solution mapping, rigorously ensuring existence, uniqueness, and global stability preservation. This work significantly extends the modeling applicability of mean-field G-SDEs in environments characterized by Knightian uncertainty.
0 citationsRead paper
Bi-SCORE for Weighted Bipartite Networks with Application in Knowledge Source Discovery
Community detection in weighted bipartite citation networks—particularly for modeling knowledge flow and core research structures in statistics—remains challenging due to citation asymmetry, degree heterogeneity, and network sparsity. Method: We propose Bi-SCORE, a spectral clustering method grounded in a weighted bipartite degree-corrected stochastic block model (DC-SBM), employing the ratio-of-eigenvectors approach. It inherently preserves citation directionality, requires no initialization, enjoys theoretical consistency guarantees, and is robust to degree heterogeneity and sparsity. Contribution/Results: Applied to a 2001–2023 citation network comprising eight leading statistics journals, Bi-SCORE identifies six coherent research domains—applied statistics, methodology, theoretical statistics, computational statistics, econometrics, and biostatistics—while accurately mapping cross-journal knowledge sources and diffusion pathways. The method delivers an interpretable, reproducible, and statistically principled framework for bipartite community detection in scholarly citation networks.
0 citationsRead paper
Recent publications
Latest Papers
Debiasing as a Measurement Intervention: Calibrated Ties and Resolution Loss in LLM-as-a-Judge Evaluation
研究通过TraceJudgeBench基准测试评估LLM作为评判时去除引用偏见的方法，发现强去偏指令虽然能减少偏见但会损害分辨能力，并提出了一种分离技术以恢复分辨率。
0 citationsRead paper
Can Large Language Models Explain Flight Safety Events? A Prior-Guided Semantic LLM-based Approach
研究提出FlightLLM，一种基于先验引导的语义大语言模型方法，通过特征工程、语义离散化及少量样本对比学习等手段解决飞行安全事件解释问题。
0 citationsRead paper
StatCite: A Large-scale Citation Network Dataset for Statistics and Data Science
This study addresses the notable scarcity of large-scale, structurally deep, and metadata-rich citation network datasets in statistics and data science. To bridge this gap, the authors construct a comprehensive citation network dataset comprising 189,101 papers published between 1981 and 2025, enriched with extensive metadata including titles, authors, abstracts, keywords, and references. For the first time in this domain, the work simultaneously integrates four complementary network types: paper citation, co-citation, bibliographic coupling, and journal citation networks. Through systematic data collection, network construction, and community detection methodologies, the dataset successfully reproduces canonical structural properties of citation networks and reveals multiple core research themes. This resource provides a high-quality foundation for advancing bibliometric analysis, knowledge graph development, and science of science research.
0 citationsRead paper
Quantitative Stability and Contraction Principles for Mean-Field G-SDEs
This study establishes a quantitative stability theory for mean-field stochastic differential equations driven by G-Brownian motion (mean-field G-SDEs) under non-Lipschitz coefficients, volatility uncertainty, and square-integrable initial data. Methodologically, we develop a novel Bihari–Osgood-type inequality within the G-expectation framework, integrating nonlinear expectation theory with mean-field analysis to derive explicit stability moduli with respect to initial conditions and coefficient perturbations; we further identify and formalize a short-time contraction property. Key contributions include: (1) the first precise quantification of solution-map sensitivity under coupled non-Lipschitz and volatility-uncertain dynamics; (2) a global stability propagation mechanism; and (3) sharp Hölder continuity estimates for the data-to-solution mapping, rigorously ensuring existence, uniqueness, and global stability preservation. This work significantly extends the modeling applicability of mean-field G-SDEs in environments characterized by Knightian uncertainty.
0 citationsRead paper
Bi-SCORE for Weighted Bipartite Networks with Application in Knowledge Source Discovery
Community detection in weighted bipartite citation networks—particularly for modeling knowledge flow and core research structures in statistics—remains challenging due to citation asymmetry, degree heterogeneity, and network sparsity. Method: We propose Bi-SCORE, a spectral clustering method grounded in a weighted bipartite degree-corrected stochastic block model (DC-SBM), employing the ratio-of-eigenvectors approach. It inherently preserves citation directionality, requires no initialization, enjoys theoretical consistency guarantees, and is robust to degree heterogeneity and sparsity. Contribution/Results: Applied to a 2001–2023 citation network comprising eight leading statistics journals, Bi-SCORE identifies six coherent research domains—applied statistics, methodology, theoretical statistics, computational statistics, econometrics, and biostatistics—while accurately mapping cross-journal knowledge sources and diffusion pathways. The method delivers an interpretable, reproducible, and statistically principled framework for bipartite community detection in scholarly citation networks.
0 citationsRead paper