About

Jennifer Pan is a political scientist whose research focuses on political communication, digital media, and authoritarian politics. She is the Sir Robert Ho Tung Professor of Chinese Studies, Professor of Communication, and a Senior Fellow at the Freeman Spogli Institute for International Studies. Additionally, she holds courtesy appointments in Political Science and Sociology. Dr. Pan's research employs experimental and computational methods with large-scale datasets on political activity to explore the role of digital media in politics, including how political censorship, propaganda, and information manipulation operate in the digital age, and how these influence political preferences and behaviors. Her work has been published in peer-reviewed journals such as the American Political Science Review, American Journal of Political Science, Journal of Politics, Science, and Nature.

Research topics

• Computer Science
• Biology
• Cell biology
• Biochemistry
• Chemistry
• Medicine
• Artificial Intelligence
• Neuroscience
• Cancer research
• Operating system
• Real-time computing
• Mathematical optimization
• Computer network
• Telecommunications
• Mathematics
• Microeconomics
• Statistics
• Immunology
• Molecular biology

Selected publications

• Aligning Multimodal Sequential Recommendations via Robust Direct Preference Optimization with Sparse MoE

  arXiv (Cornell University) · 2026-03-31

  articleOpen accessSenior author

  Preference-based alignment objectives have been widely adopted, from RLHF-style pairwise learning in large language models to emerging applications in recommender systems. Yet, existing work rarely examines how Direct Preference Optimization (DPO) behaves under implicit feedback, where unobserved items are not reliable negatives. We conduct systematic experiments on multimodal sequential recommendation to compare common negative-selection strategies and their interaction with DPO training. Our central finding is that a simple modification, replacing deterministic hard negatives with stochastic sampling from a dynamic top-K candidate pool, consistently improves ranking performance. We attribute its effectiveness to two factors: (1) reducing erroneous suppressive gradients caused by false negatives, and (2) retaining informative hard signals while smoothing optimization via controlled stochasticity. With an optional sparse Mixture-of-Experts encoder for efficient capacity scaling, RoDPO achieves up to 5.25% NDCG@5 on three Amazon benchmarks, with nearly unchanged inference cost.

  PublisherOA PDF

• Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models

  ArXiv.org · 2026-03-16

  articleOpen accessSenior author

  Large Language Models (LLMs) have demonstrated impressive capabilities in simulating diverse human behaviors and personalities. However, existing methods for personality control, which include prompt engineering and standard Supervised Fine-Tuning (SFT), typically treat personality traits as discrete categories (e.g., "Extroverted" vs. "Introverted"), lacking the ability to precisely control the intensity of a trait on a continuous spectrum. In this paper, we introduce Fusian, a novel framework for fine-grained, continuous personality control in LLMs. Fusian operates in two stages: (1) Trajectory Collection, where we capture the dynamic evolution of personality adoption during SFT by saving a sequence of LoRA adapters, effectively mapping the continuous manifold of a trait; and (2) RL-based Dynamic Fusion, where we train a policy network using Reinforcement Learning to dynamically compute mixing weights for these frozen adapters. By sampling from a Dirichlet distribution parameterized by the policy network, Fusian fuses multiple adapters to align the model's output with a specific numerical target intensity. Experiments on the Qwen3-14B model demonstrate that Fusian achieves high precision in personality control, significantly outperforming baseline methods in aligning with user-specified trait intensities.

  PublisherOA PDF

• Dynamic Fault Tree Analysis using Phase-type Distribution and Probabilistic Model Checking

  2026-01-26

  article1st authorCorresponding

  SUMMARY & CONCLUSIONSThis paper presents an advanced methodological framework for reliability analysis by integrating dynamic fault tree (DFT) modeling with phase-type (PH) distributions, further enhanced by probabilistic model checking techniques. DFTs are widely employed for representing complex system behaviors, such as functional dependencies and order-dependent failures, that standard static fault trees cannot adequately capture. While DFTs provide sophisticated modeling capabilities, translating them into explicit probabilistic representations for precise failure time analysis remains challenging. Phase-type distributions offer a compelling solution to this challenge. Defined by the absorption times of finite-state continuous-time Markov chain (CTMC) processes, PH distributions can approximate any positive distribution arbitrarily closely. This paper explicitly demonstrates how the failure time distributions derived from the DFT inherently conform to PH distributions. This theoretical connection provides a rigorous and systematic approach to mapping various DFTs into their corresponding PH distribution models. In particular, the link between DFT and PH distributions enhances the interpretability of the system failure distribution with DFT basic events. The new framework enables the use of arbitrary PH distributions to approximate any system lifetime distribution. Furthermore, the efficient translation from DFT to CTMC enables rigorous, automated probabilistic model checking, which significantly enhances the analytical power and efficiency of DFT analysis by allowing for the precise computation of reliability metrics, such as system unreliability, mean-time-to-failure (MTTF), and other critical dependability measures.

  PublisherDOI

• Aligning Multimodal Sequential Recommendations via Robust Direct Preference Optimization with Sparse MoE

  arXiv (Cornell University) · 2026-03-31

  preprintOpen accessSenior author

  Preference-based alignment objectives have been widely adopted, from RLHF-style pairwise learning in large language models to emerging applications in recommender systems. Yet, existing work rarely examines how Direct Preference Optimization (DPO) behaves under implicit feedback, where unobserved items are not reliable negatives. We conduct systematic experiments on multimodal sequential recommendation to compare common negative-selection strategies and their interaction with DPO training. Our central finding is that a simple modification, replacing deterministic hard negatives with stochastic sampling from a dynamic top-K candidate pool, consistently improves ranking performance. We attribute its effectiveness to two factors: (1) reducing erroneous suppressive gradients caused by false negatives, and (2) retaining informative hard signals while smoothing optimization via controlled stochasticity. With an optional sparse Mixture-of-Experts encoder for efficient capacity scaling, RoDPO achieves up to 5.25% NDCG@5 on three Amazon benchmarks, with nearly unchanged inference cost.

  PublisherDOI

• Competition in the Segregation Mechanism of Granular Flow Within a 2D Rotating Drum Based on Magnetic Positioning Technology

  Sensors · 2026-03-10

  articleOpen access1st author

  Accurate monitoring of internal particle motion in dense granular flows remains a significant challenge across various fields, ranging from geophysics to industrial processes. To address the limitations of existing observational techniques, this study presents a novel high-precision magnetic array positioning system based on magnetic dipole theory for dynamically tracking individual particles within opaque granular media. The system integrates an array of nine magnetic sensors with a hybrid optimization algorithm that combines Particle Swarm Optimization (PSO) and gradient-based local refinement, achieving a dynamic positioning accuracy within the maximum measurable range, with a maximum dynamic error of 2.5 ± 0.5 mm and a trajectory continuity exceeding 99%. Deployed in a quasi-two-dimensional rotating drum, the system enables detailed investigation of particle segregation mechanisms. Reconstruction and analysis of the trajectories of a high-density intruder (magnetic bead) allow quantification of the competition among segregation mechanisms through the Froude number. The results reveal three distinct motion phases with increasing rotational speed: a gravity-dominated percolation stage, a transitional collision-diffusion competition stage, and a centrifugal diffusion-dominated stage. Each phase exhibits unique kinematic signatures governed by the interplay of inertial, gravitational, and contact forces. This work not only establishes a robust and accurate sensor-based method for internal granular flow monitoring but also provides new mechanistic insights into segregation dynamics, with implications for understanding geological hazards such as debris flows.

  PublisherOA PDFDOI

• From statistical modeling to AI-integrated inverse Gaussian process: A comprehensive review for prognostics and health management

  ENGINEERING Management · 2026-03-01

  articleOpen access

  Abstract In prognostics and health management (PHM), degradation modeling plays a central role in reliability analysis and lifetime prediction. The inverse Gaussian (IG) process has recently attracted increasing attention for its ability to describe monotonic and cumulative degradation with heavy-tailed behavior, analytical tractability, and clear physical interpretability. Meanwhile, the rapid development of artificial intelligence (AI) has created new opportunities to combine statistical modeling with learning-based approaches in reliability analysis. This paper presents a comprehensive review of IG-process-based degradation modeling, covering its theoretical foundations, model extensions, parameter estimation, and diagnostic methods. Applications in accelerated degradation test design, burn-in test, remaining useful life prediction, and maintenance optimization are systematically summarized. Recent progress on AI-integrated IG frameworks is also reviewed and critically assessed. In addition, key challenges and research opportunities are discussed to guide future developments in intelligent PHM.

  PublisherOA PDFDOI

• Prospect-Theory-Based Modeling of GENCO Strategies in Electricity–Carbon Coupled Markets

  2026-01-18

  article

  Driven by carbon neutrality goals and carbon pricing reforms, power generation companies (GENCOs) are facing increasingly complex decision-making challenges. This paper first develops a bilevel modeling framework to describe the participation of multiple GENCOs in the electricity–carbon market. Then, the upper-level model is established to capture the optimal decision-making behavior of GENCOs based on prospect theory (PT) theory, and the lower-level model represents the day-ahead (DA) market clearing process. The Karush-Kuhn-Tucker (KKT) conditions are applied to reformulate the bilevel structure into a single-level MPEC. Finally, based on a modified IEEE 30-bus system, the equilibrium of the electricity–carbon market is solved and GENCO strategies are analyzed. Simulation results show that increasing carbon preference significantly reduces system emissions and raises carbon prices, while high-emission enterprises exhibit more pronounced strategic shifts, validating the effectiveness of the proposed behavioral model.

  PublisherDOI

• Bridging Heritage Knowledge and Digital Models: An HBIM Integration Framework

  ISPRS annals of the photogrammetry, remote sensing and spatial information sciences · 2025-09-24

  articleOpen accessSenior author

  Abstract. Architectural heritage conservation demands the integration of precise physical documentation and interpretative design knowledge, yet current HBIM approaches remain fragmented: ‘scan-to-BIM’ prioritizes geometric accuracy at the expense of semantic richness, while “rule-based reconstruction” emphasizes idealized logic over as-built evidence. To bridge this gap, this study introduces the KSQI paradigm (Knowledge-Semantics-Quantities-Image), a novel framework that systematically connects domain expertise with digital modelling to balance spatial accuracy and architectural semantics. The research develops an as-recognized modelling or semantic-driven modelling through (1) a conservation cycle-guided information indexing system for semantic-driven knowledge integration, (2) a data-model decoupling workflow that teams from different disciplines maintain their working habits, handling data and models separately, then recoupling data-model by BIM team, and (3) a pattern book tooling solution including check forms for hierarchical investigation, algorithm modelling generator. By linking physical attributes (quantities/images) with design logic (semantics/knowledge), KSQI enhances information management, supports iterative knowledge updates, and facilitates informed conservation decisions. Case studies demonstrate its effectiveness in encoding both as-built conditions and historical, traditional design/construction principles, reinforcing the ‘H’ (history/heritage knowledge) in HBIM. This framework advances heritage documentation toward the smart metric survey, ensuring models serve as dynamic, semantically rich assets for conservation, research, dissemination, and digital twin applications.

  PublisherOA PDFDOI

• A multiple-feature-based recognition algorithm for OFDM signals and subcarrier modulations

  IET conference proceedings. · 2025-10-13

  article1st authorCorresponding

  Aiming at the problem of low accuracy of traditional methods in the recognition of OFDM signals and their subcarrier modulation modes, a multi-feature based OFDM signal and subcarrier modulation recognition algorithm is proposed. By extracting the fourth-order cumulant, sixth-order cumulant, and constellation diagram features of the signal, featuresHs1,Hs2, andHs3are constructed. Using the fourth-order cumulant and the first two features, the recognition of OFDM, 8PSK, BPSK, and QPSK signals is realized. An improved K-means++ clustering algorithm is further designed to determine the clustering centroids of the constellation diagrams and complete the classification and identification of 16QAM, 32QAM, 64QAM, 128QAM and 256QAM signals based on the constructed features. The simulation results show that the algorithm has good recognition performance, and improves the recognition accuracy by 15% at SNR=5dB compared with the traditional algorithm for high-order MQAM signals, which can be effectively applied to the modulation recognition of OFDM signals and their subcarrier signals.

  PublisherDOI

• Interpretable and Sample-Efficient Reinforcement Learning by Discovering Value-Based HTN Models

  2025-04-25

  articleSenior author

  Despite great success across various domains, reinforcement learning (RL) faces significant challenges in sample efficiency and interpretability, particularly in sparse reward scenarios. To address these issues, researchers have explored integrating symbolic planning with RL, leveraging symbolic models to provide high-level guidance to RL agents. However, existing methods often rely on manually crafted knowledge or overlook hierarchical task structures, leading to suboptimal planning, ineffective option policy training and low sampling efficiency in complex environments. In this paper, we thus propose a novel RL framework that incorporates value-based hierarchical symbolic models learned directly from the environment. This framework establishes a loop training process to automatically construct value-based hierarchical task network (HTN) models, including action models and method models embedded with intrinsic value, derived from RL interaction trajectories. The learned hierarchical symbolic models alleviate the reliance on expert domain knowledge, provide interpretable and globally optimal plans to guide the training of low-level option policies, and significantly enhance sample efficiency. We empirically validate the interpretability of subtasks and demonstrate improved sample efficiency compared to state-of-the-art approaches across several domains.

  PublisherDOI

Recent grants

• Collaborative Research: Quantitative Reliability Prediction in Early Design Stages

  NSF · $245k · 2013–2017

• Collaborative Research: Efficient Experimentation for Product and Process Reliability Improvement

  NSF · $348k · 2009–2013

• Design of Experiments with Dynamic Responses

  NSF · $315k · 2017–2021

• Modeling and Analysis of Profiled Reliability Tests Using Computation-Intensive Statistical Methods

  NSF · $263k · 2006–2010

Frequent coauthors

• Ke Jian Liu

  Stony Brook School

  44 shared

• Qiang Yang

  44 shared

• Xiuyun Zhu

  Nuclear and Radiation Safety Center

  39 shared

• Zhifeng Qi

  Capital Medical University

  33 shared

• Xunming Ji

  Chinese Institute for Brain Research

  30 shared

• Yongmei Zhao

  Capital Medical University

  29 shared

• Rongqiao He

  Guangzhou Experimental Station

  24 shared

• Yumin Luo

  Chinese Institute for Brain Research

  19 shared

Awards & honors

• John S. Knight Journalism Fellowships