Selected work

Representative Papers

Empirical evidence of Large Language Model's influence on human spoken communication

Sep 03, 2024arXiv.org

This study investigates whether large language models (LLMs), such as ChatGPT, reshape human spoken language and cultural practices through human–AI linguistic feedback loops. Method: Leveraging ASR-transcribed speech from 280,000 university-level YouTube lecture videos, we conduct time-series word frequency analysis and employ a pre–post ChatGPT release quasi-experimental design. Contribution/Results: We present the first empirical evidence that LLMs directly influence authentic human spoken behavior: post-release, ChatGPT-characteristic lexical items exhibit statistically significant increases in academic speech (p < 0.001), confirming systematic oral imitation by humans. Moving beyond prior written-language–focused work, this study reveals the mechanism of AI-generated language diffusion into spoken discourse. It further highlights critical sociocultural risks—including erosion of linguistic diversity, discursive manipulation, and asymmetric human–AI co-evolution—thereby advancing foundational understanding of LLMs’ real-world linguistic impact.

11 citations1 influentialRead paper

A Comprehensive Study on the Impact of Vulnerable Dependencies on Open-Source Software

Oct 28, 2024IEEE International Symposium on Software Reliability Engineering

This study systematically investigates the severity, persistence, and distribution patterns of dependency vulnerabilities in open-source projects, and their correlations with project-scale and activity metrics. Method: Leveraging our custom-built Software Composition Analysis (SCA) tool, VODA, we constructed a high-quality, multi-language dataset encompassing 1,000+ open-source projects and over 50,000 version releases. Contribution/Results: We quantitatively reveal—first time at scale—that critical vulnerabilities predominantly originate from transitive dependencies and exhibit median remediation latency exceeding 12 months; team size correlates significantly with faster vulnerability resolution; and vulnerability persistence and propagation patterns exhibit cross-language commonalities. Based on these findings, we propose the “high-risk-dependency-first governance” principle. The study provides empirical evidence and methodological support for optimizing dependency update mechanisms and strengthening software supply chain security governance.

5 citationsRead paper

Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models

Jan 20, 2026

This work proposes a practical, three-stage “Locate–Guide–Improve” framework that transforms mechanistic interpretability from a post-hoc diagnostic tool into an engineering-driven optimization methodology for large language models. By systematically integrating techniques for identifying critical neurons and pathways with targeted interventions—such as activation manipulation and module editing—the framework establishes a standardized protocol for model refinement while clearly distinguishing between localization and guidance mechanisms. Empirical results demonstrate significant improvements in model alignment, task performance, and reasoning efficiency, thereby advancing mechanistic interpretability toward real-world applicability.

4 citationsRead paper

Foundations for an Abstract Proof Theory in the Context of Horn Rules

Apr 12, 2023arXiv.org

Horn logic lacks a unified proof-theoretic framework. Method: We introduce an abstract, logic-agnostic proof-theoretic model based on *g-sequents* (generalized binary graph-like sequents), formalizing them via abstract algebraic semantics, graph representation, and lattice-theoretic modeling; we develop an algebraic classification and interaction analysis theory for inference rules, and design a generic proof transformation algorithm that characterizes upper and lower bound correspondences of nested/marked sequent systems within abstract lattices. Contribution/Results: We achieve the first uniform characterization of diverse Horn-logic sequent-style systems; establish that cut-free nested/marked systems precisely correspond to extremal points (i.e., top and bottom elements) of the associated lattice; derive quantitative relationships between proof size and sequent complexity; and determine the computational complexity of lattice operations. This work bridges structural proof theory, algebraic logic, and lattice theory to provide a foundational, modular framework for Horn-logic reasoning.

3 citationsRead paper

An Overview of Arithmetic Adaptations for Inference of Convolutional Neural Networks on Re-configurable Hardware

May 19, 2025

Deploying CNNs—particularly TinyYOLOv3—on resource-constrained FPGAs (e.g., Xilinx Artix-7) faces fundamental trade-offs among computational demand, memory footprint, and hardware arithmetic constraints. To address this, we propose an end-to-end lightweight adaptation framework integrating batch normalization fusion, structured filter pruning, and post-training INT8 quantization, coupled with hardware-aware operator mapping and Vivado HLS co-optimization. Our approach significantly improves energy efficiency and hardware resource utilization: on Artix-7, it achieves end-to-end real-time object detection with 3.1× higher inference throughput, 42% lower power consumption, and 37% reduction in LUT usage. The methodology establishes a highly efficient, reproducible FPGA implementation paradigm for edge-deployed CNNs.

2 citationsRead paper

Recent publications

Latest Papers

The Sync Heap: Delete First, Ask Questions Later

Aug 07, 2026

This work challenges the classical logarithmic lower bound for insertion and extract-min operations in comparison-based heaps, which stems from the requirement to immediately identify the element being deleted. Building upon Chazelle’s soft heap framework, the paper introduces a novel deterministic data structure—termed the *sync heap*—that decouples deletion from the immediate revelation of element identity. By incorporating mechanisms such as delayed synchronization of deletions, operation reordering, and compression, the sync heap achieves constant amortized time complexity for both insertions and deletions under the practical assumption that the user performs only a constant number of queries. This breakthrough not only surpasses the long-standing logarithmic barrier but also improves the optimal solution for the unit-time scheduling problem from $O(n \log n)$ to $O(n)$.

0 citationsRead paper

Example-Guided Prompting for Document-Level Text Simplification

Aug 05, 2026

Large language models (LLMs) often struggle to simultaneously preserve semantic content, ensure readability, and maintain discourse coherence in document-level text simplification when relying solely on instruction-based prompting. To address this limitation, this work proposes a retrieval-augmented, exemplar-guided prompting approach that dynamically retrieves relevant examples from a parallel simplification corpus and incorporates them into the prompt, thereby guiding the model to produce more consistent and higher-quality simplified texts without requiring task-specific fine-tuning. Evaluated systematically on the OneStopEnglish corpus, the proposed method significantly outperforms pure prompting baselines and matches or exceeds the performance of supervised and planning-based systems such as T5 and PlanSimp. Furthermore, this study provides the first empirical analysis of the varying capacities among different LLMs to effectively leverage retrieved exemplars for simplification.

0 citationsRead paper

Self-Healing 6G Networks-in-Network for Resilient Wireless Communication

Aug 05, 2026

This work addresses the vulnerability of 6G networks employing multiple autonomous subnetworks sharing spectrum, which are prone to communication disruptions caused by failures or malicious interference, necessitating efficient self-healing mechanisms. The paper proposes a novel Nested-in-Network (NiN) architecture that, for the first time, integrates spectrum monitoring and dynamic resource management into a unified control loop, enabling end-to-end autonomous response from anomaly detection to frequency reconfiguration. By incorporating continuous spectrum scanning, lightweight anomaly detection algorithms, and automated subnetwork reconfiguration, the proposed architecture significantly enhances system resilience and recovery speed under interference, as demonstrated in real-world experiments.

0 citationsRead paper

The Push-Forward Transform for Continuous and Robust Comparison of Dynamic Shapes

Aug 03, 2026

This work addresses the challenge of continuously, robustly, and interpretably comparing dynamic shapes and their temporal evolution while preserving invariance under transformations such as translation, rotation, reflection, reparameterization, and uniform scaling. To this end, the authors propose a novel framework based on Push-Forward mappings—introduced here for the first time in shape analysis—that transports signed distance functions (SDFs) onto a common reference domain, yielding a continuous invariant representation that simultaneously encodes both boundary and interior geometric information. The approach unifies the treatment of 2D, 3D, and dynamic shapes, supports joint analysis with associated scalar fields, and enables an interpretable shape metric capable of revealing structural features such as skeletal topology and rotational symmetries. Extensive experiments demonstrate the method’s effectiveness and robustness across diverse datasets.

0 citationsRead paper

Beyond Lanes: Traffic Flow Dynamics in Disordered Conditions Based on High-Resolution Trajectory Data

Aug 01, 2026

This study addresses the limitations of conventional lane-based traffic flow models in capturing chaotic traffic conditions characterized by the absence of marked lanes, high heterogeneity, and persistent lateral interactions. Leveraging high-resolution drone trajectory data, it integrates a two-dimensional Edie macroscopic framework with microscopic car-following analysis to construct, for the first time, a two-dimensional macroscopic fundamental diagram tailored to unlaned mixed traffic. The work reveals the critical role of lateral redistribution in congestion propagation and quantifies temporal headways and lateral spacing distributions across vehicle types through spatiotemporal speed fields, identification of steady-state car-following relationships, and explicit modeling of vehicular heterogeneity. These advances enable accurate characterization of stop-and-go wave dynamics and provide a data-driven foundation for calibrating and validating unlaned traffic flow models.

0 citationsRead paper