Selected work

Representative Papers

An advancing-ridge approach for recovering boundary $(d-1)$-simplices in $d$-dimensional meshes

Aug 15, 2026

This study addresses the challenge of high-dimensional boundary-constrained mesh recovery in four-dimensional spacetime simulations by proposing an advancing front algorithm based on (d−2)-simplex ridges. The method innovatively integrates a constrained cavity operator with an incremental Steiner point insertion strategy to effectively ensure geometric consistency and topological correctness of complex boundaries. Experimental results demonstrate that the algorithm achieves a boundary recovery rate exceeding 99% in four-dimensional scenarios, generating 300 million pentatopes within merely 15 minutes. These findings indicate a significant breakthrough in overcoming efficiency and robustness bottlenecks associated with high-dimensional constrained mesh generation, thereby providing reliable computational support for large-scale spacetime numerical simulations.

0 citationsRead paper

Growing Hypergraphs with Homophily

Jul 17, 2026

This work addresses the limitation of existing hypergraph models that commonly assume conditional independence among hyperedges, which fails to capture the intricate dependencies between hyperedges and node labels as well as homophily mechanisms in real-world higher-order systems. We propose a growing dynamic hypergraph generative model that introduces inter-hyperedge dependencies through noisy copying of existing hyperedges and incorporates a multi-label-driven homophily mechanism to produce tunable assortative structures. By relaxing the hyperedge independence assumption, our model offers theoretical interpretability and supports likelihood-based parameter inference and community detection. We develop a stochastic EM algorithm for parameter estimation, integrate simulated annealing for community discovery, and analytically characterize the power-law degree distribution and the asymptotic behavior of intra-hyperedge label joint distributions. Experiments demonstrate that the model effectively captures homophily on both synthetic and real datasets, significantly outperforming edge-independent approaches in community detection.

0 citationsRead paper

Experimental Evidence on the Learning Impact of Generative AI

Jul 09, 2026

This study investigates the short- and long-term effects of generative AI on students’ knowledge acquisition and higher-order writing skills. In a randomized controlled experiment, undergraduate participants learned a novel topic and composed analytical essays either with or without AI assistance, followed by unaided assessments immediately and one week later. Results indicate that AI support improved immediate knowledge test performance by 0.27 standard deviations, with this advantage persisting after one week. Further analysis distinguishing between “augmentation” and “automation” usage patterns revealed that augmentation—where AI complements rather than replaces student effort—significantly enhanced writing quality in subsequent unaided tasks, alongside increased information-seeking behaviors and greater reported enjoyment of learning. The findings highlight a delayed benefit of generative AI in educational contexts and underscore the critical moderating role of how such tools are employed.

0 citationsRead paper

Dark&Stormy: Modeling Humor in the Worst Sentences Ever Written

Oct 28, 2025

This study addresses the challenge of detecting deliberately crafted “bad humor” in English—a genre where state-of-the-art humor detection models exhibit significant performance degradation. To tackle this, we introduce the first bad-humor corpus derived from the Bulwer-Lytton Fiction Contest, systematically analyzing its structural patterns involving puns, irony, metaphor, and metafictional devices. We conduct the first human–LLM comparative study on bad-humor generation, revealing that LLMs over-rely on specific rhetorical devices and nonce collocations, exposing a rhetorical control bias. Integrating literary rhetorical analysis, controllable prompt engineering, and human–AI collaborative evaluation, we demonstrate that current models lack robust semantic–stylistic disentanglement capabilities for low-quality humor. Our contributions include: (1) a novel, manually annotated bad-humor benchmark; (2) empirical evidence of LLMs’ rhetorical limitations; and (3) open-sourced data and code to advance computational humor research.

0 citationsRead paper

Developing Training Procedures for Piecewise-linear Spline Activation Functions in Neural Networks

Sep 16, 2025

This work addresses the limitation of fixed, non-adaptive activation functions in neural networks. We propose a learnable piecewise-linear B-spline activation function, jointly optimized end-to-end with network weights to enable dynamic adaptation to task-specific data distributions. Methodologically, we design and systematically evaluate nine bi-level optimization training strategies. Extensive experiments on feedforward neural networks (FNNs) and convolutional neural networks (CNNs) demonstrate substantial improvements: up to 94% reduction in test error for FNNs and 51% for CNNs. The approach significantly enhances both parameter efficiency and predictive accuracy. However, these gains come at the cost of moderately increased training complexity and marginal inference latency. Our study establishes a novel paradigm for structured, learnable activation function modeling and provides an empirical benchmark for future research in adaptive activation design.

0 citationsRead paper

Recent publications

Latest Papers

An advancing-ridge approach for recovering boundary $(d-1)$-simplices in $d$-dimensional meshes

Aug 15, 2026

This study addresses the challenge of high-dimensional boundary-constrained mesh recovery in four-dimensional spacetime simulations by proposing an advancing front algorithm based on (d−2)-simplex ridges. The method innovatively integrates a constrained cavity operator with an incremental Steiner point insertion strategy to effectively ensure geometric consistency and topological correctness of complex boundaries. Experimental results demonstrate that the algorithm achieves a boundary recovery rate exceeding 99% in four-dimensional scenarios, generating 300 million pentatopes within merely 15 minutes. These findings indicate a significant breakthrough in overcoming efficiency and robustness bottlenecks associated with high-dimensional constrained mesh generation, thereby providing reliable computational support for large-scale spacetime numerical simulations.

0 citationsRead paper

Growing Hypergraphs with Homophily

Jul 17, 2026

This work addresses the limitation of existing hypergraph models that commonly assume conditional independence among hyperedges, which fails to capture the intricate dependencies between hyperedges and node labels as well as homophily mechanisms in real-world higher-order systems. We propose a growing dynamic hypergraph generative model that introduces inter-hyperedge dependencies through noisy copying of existing hyperedges and incorporates a multi-label-driven homophily mechanism to produce tunable assortative structures. By relaxing the hyperedge independence assumption, our model offers theoretical interpretability and supports likelihood-based parameter inference and community detection. We develop a stochastic EM algorithm for parameter estimation, integrate simulated annealing for community discovery, and analytically characterize the power-law degree distribution and the asymptotic behavior of intra-hyperedge label joint distributions. Experiments demonstrate that the model effectively captures homophily on both synthetic and real datasets, significantly outperforming edge-independent approaches in community detection.

0 citationsRead paper

Experimental Evidence on the Learning Impact of Generative AI

Jul 09, 2026

This study investigates the short- and long-term effects of generative AI on students’ knowledge acquisition and higher-order writing skills. In a randomized controlled experiment, undergraduate participants learned a novel topic and composed analytical essays either with or without AI assistance, followed by unaided assessments immediately and one week later. Results indicate that AI support improved immediate knowledge test performance by 0.27 standard deviations, with this advantage persisting after one week. Further analysis distinguishing between “augmentation” and “automation” usage patterns revealed that augmentation—where AI complements rather than replaces student effort—significantly enhanced writing quality in subsequent unaided tasks, alongside increased information-seeking behaviors and greater reported enjoyment of learning. The findings highlight a delayed benefit of generative AI in educational contexts and underscore the critical moderating role of how such tools are employed.

0 citationsRead paper

Dark&Stormy: Modeling Humor in the Worst Sentences Ever Written

Oct 28, 2025

This study addresses the challenge of detecting deliberately crafted “bad humor” in English—a genre where state-of-the-art humor detection models exhibit significant performance degradation. To tackle this, we introduce the first bad-humor corpus derived from the Bulwer-Lytton Fiction Contest, systematically analyzing its structural patterns involving puns, irony, metaphor, and metafictional devices. We conduct the first human–LLM comparative study on bad-humor generation, revealing that LLMs over-rely on specific rhetorical devices and nonce collocations, exposing a rhetorical control bias. Integrating literary rhetorical analysis, controllable prompt engineering, and human–AI collaborative evaluation, we demonstrate that current models lack robust semantic–stylistic disentanglement capabilities for low-quality humor. Our contributions include: (1) a novel, manually annotated bad-humor benchmark; (2) empirical evidence of LLMs’ rhetorical limitations; and (3) open-sourced data and code to advance computational humor research.

0 citationsRead paper

Developing Training Procedures for Piecewise-linear Spline Activation Functions in Neural Networks

Sep 16, 2025

This work addresses the limitation of fixed, non-adaptive activation functions in neural networks. We propose a learnable piecewise-linear B-spline activation function, jointly optimized end-to-end with network weights to enable dynamic adaptation to task-specific data distributions. Methodologically, we design and systematically evaluate nine bi-level optimization training strategies. Extensive experiments on feedforward neural networks (FNNs) and convolutional neural networks (CNNs) demonstrate substantial improvements: up to 94% reduction in test error for FNNs and 51% for CNNs. The approach significantly enhances both parameter efficiency and predictive accuracy. However, these gains come at the cost of moderately increased training complexity and marginal inference latency. Our study establishes a novel paradigm for structured, learnable activation function modeling and provides an empirical benchmark for future research in adaptive activation design.

0 citationsRead paper