About

Lu Liu is an Assistant Professor in the School of Modern Languages at Georgia Tech. Her research broadly explores the intersections of literary, visual, and media cultures with the transnational production of biological, hygienic, and environmental knowledge. She is completing a book manuscript titled "Trans-species Revolution: Pests and the Unwanted Nature in Modern China, 1900s–1970s," which examines the critical role of pests as material and symbolic agents in shaping political struggles, enlightenment movements, socialist construction, and ecological engineering in Chinese history. Her work argues that the dual aspirations of the Chinese revolution—building a new world and eradicating unwanted nature—are profoundly intertwined and require critical examination of each other. Her research has appeared or is forthcoming in publications such as China Perspectives, the British Journal of the History of Science, and edited volumes on global humanities and East Asian media. Her teaching interests include Chinese literature, culture, and cinema, Chinese and Asian American foodways, revolutionary art in the Mao era, and sustainability in contemporary Chinese society.

Research topics

• Artificial Intelligence
• Machine Learning
• Computer Science
• Computer Security
• Data Mining
• Cognitive psychology
• Psychology

Selected publications

• Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable

  ArXiv.org · 2025-03-01

  preprintOpen accessSenior author

  Safety alignment is an important procedure before the official deployment of a Large Language Model (LLM). While safety alignment has been extensively studied for LLM, there is still a large research gap for Large Reasoning Models (LRMs) that equip with improved reasoning capability. We in this paper systematically examine a simplified pipeline for producing safety aligned LRMs. With our evaluation of various LRMs, we deliver two main findings: i) Safety alignment can be done upon the LRM to restore its safety capability. ii) Safety alignment leads to a degradation of the reasoning capability of LRMs. The two findings show that there exists a trade-off between reasoning and safety capability with the sequential LRM production pipeline. The discovered trade-off, which we name Safety Tax, should shed light on future endeavors of safety research on LRMs. As a by-product, we curate a dataset called DirectRefusal, which might serve as an alternative dataset for safety alignment. Our source code is available at https://github.com/git-disl/Safety-Tax.

  PublisherOA PDFDOI

• MolLangBench: A Comprehensive Benchmark for Language-Prompted Molecular Structure Recognition, Editing, and Generation

  arXiv (Cornell University) · 2025-05-21 · 1 citations

  preprintOpen access

  Precise recognition, editing, and generation of molecules are essential prerequisites for both chemists and AI systems tackling various chemical tasks. We present MolLangBench, a comprehensive benchmark designed to evaluate fundamental molecule-language interface tasks: language-prompted molecular structure recognition, editing, and generation. To ensure high-quality, unambiguous, and deterministic outputs, we construct the recognition tasks using automated cheminformatics tools, and curate editing and generation tasks through rigorous expert annotation and validation. MolLangBench supports the evaluation of models that interface language with different molecular representations, including linear strings, molecular images, and molecular graphs. Evaluations of state-of-the-art models reveal significant limitations: the strongest model (GPT-5) achieves $86.2\%$ and $85.5\%$ accuracy on recognition and editing tasks, which are intuitively simple for humans, and performs even worse on the generation task, reaching only $43.0\%$ accuracy. These results highlight the shortcomings of current AI systems in handling even preliminary molecular recognition and manipulation tasks. We hope MolLangBench will catalyze further research toward more effective and reliable AI systems for chemical applications.The dataset and code can be accessed at https://huggingface.co/datasets/ChemFM/MolLangBench and https://github.com/TheLuoFengLab/MolLangBench, respectively.

  PublisherOA PDFDOI

• WD-PSTALSTM: a data-driven hybrid model for prediction of diesel vehicle NOx emissions

  Energy and AI · 2025-07-26 · 2 citations

  articleOpen access1st author

  • WD-PSTALSTM combines wavelet decomposition and spatiotemporal attention for NOx prediction. • Wavelet decomposition reduces data non-stationarity, improving prediction accuracy. • Parallel spatiotemporal attention boosts model’s predictive performance on NOx emissions. • WD-PSTALSTM outperforms SVR, LSTM, and Bi-LSTM, improving MAE by 20% and RMSE by 15%. • Model shows strong potential for accurate emission inventories and pollution control. Accurate prediction of transient nitrogen oxides (NOx) emissions from diesel vehicles is essential for precise emission inventories and effective pollution control but challenged by data nonlinearity and dynamic operating conditions. This study develops the Wavelet Decomposition (WD)-Parallel Spatio-Temporal Attention-based Long Short-Term Memory (PSTALSTM) model, using real-world Portable Emission Measurement System (PEMS) and On-Board Diagnostics (OBD) data. WD preprocessing reduces emission data non-stationarity, generating more stable inputs. The PSTALSTM architecture, built upon Bidirectional Long Short-Term Memory (Bi-LSTM), incorporates a parallel attention mechanism to adaptively weight features and temporal segments, effectively capturing spatio-temporal correlations within the emission data. Validation with on-road test data demonstrates WD-PSTALSTM's superior performance over existing models. It achieves reductions exceeding 20% in mean absolute error (MAE) and 15% in root mean square error (RMSE), significantly enhancing prediction accuracy. These results establish WD-PSTALSTM as an effective approach for forecasting transient diesel engine NOx emissions. The research provides valuable methodologies for emission prediction based on vehicle operational data, contributing to environmental pollution mitigation efforts.

  PublisherDOI

• Two-Dimensional DOA Estimation of Underwater Acoustic Signals With Gain-Phase Errors

  IEEE Transactions on Communications · 2025-07-14 · 2 citations

  articleSenior author

  The intrinsic qualities of the underwater acoustic channel (UAC), such as multipath propagation and motion-induced Doppler shift, may result in large direction of arrival (DOA) estimation bias. In this paper, we propose a two-dimensional DOA estimation method for the underwater acoustic signals with gain-phase errors, coined as ML-ANM (Maximum Likelihood-Atomic Norm Minimization). We assume that the number of received signals satisfies the parameter identifiability condition. First, we propose a bimodal atomic norm involving noncoherent and coherent signals under the maximum likelihood estimation (MLE) framework, which allows simultaneous processing of noncoherent and coherent signals without decorrelation. To avoid the direct calculation of gain-phase errors, we transform the primal MLE optimization model into an ML-ANM optimization problem with two positive semidefinite constraints. Second, we utilize oblique projection to divide the optimal problem into two optimal subproblems involving noncoherent and coherent signals, and prove the equivalence of the problems and their solutions. Then, we obtain the global optimal solutions for the two optimal subproblems. Third but not the least, we provide the Cramér-Rao lower bound (CRB) and prove that the asymptotic error covariance matrix of ML-ANM achieves the CRB. Extensive experiments through realistic ocean simulations demonstrate the effectiveness of the ML-ANM method.

  PublisherDOI

• Mechanisms underpinning cold tolerance in indica hybrid rice: Insights for northern adaptation

  Gene Reports · 2025-03-28 · 1 citations

  article
  PublisherDOI

• PokéLLMon: A Grounding and Reasoning Benchmark for Large Language Models in Pokémon Battles

  ACM Transactions on Internet Technology · 2025-10-07

  articleOpen accessSenior author

  Developing grounding techniques for LLMs poses two requirements for interactive environments, i.e., (i) the presence of rich knowledge beyond the scope of existing LLMs and (ii) the complexity of tasks that require strategic reasoning. Existing environments fail to meet both requirements due to their simplicity or reliance on commonsense knowledge already encoded in LLMs for interaction. In this article, we present PokéLLMon, a new benchmark enriched with fictional game knowledge and characterized by the intense, dynamic, and adversarial gameplay of Pokémon battles, setting new challenges for the development of grounding and reasoning techniques in interactive environments. Empirical evaluations demonstrate that existing LLMs lack game knowledge and struggle in Pokémon battles. We investigate grounding techniques that leverage feedback and game knowledge, and provide a thorough analysis of reasoning methods from a new perspective of action consistency. Additionally, we introduce higher-level reasoning challenges when playing against human players. The implementation of our benchmark is released at: https://github.com/git-disl/PokeLLMon .

  PublisherDOI

• Trends and Characteristics of Occupational Radiation Exposure in the Period from 2009 to 2019 in Hubei Province, China

  Radioprotection · 2025-10-11

  article1st authorCorresponding

  This study examines occupational ionizing radiation exposure in Hubei Province, China, from 2009 to 2019, focusing on three key occupational categories (medical, industrial, and others). A total of 126,291 workers participated, with numbers increasing from 3,270 in 2009 to 20,961 in 2019. The average annual effective dose was 0.43 mSv, contributing to a total dose of 54,133.42 man·mSv over 11 years. Most workers (91.7%) received doses below the public limit of 1 mSv annually. Specifically, the proportion of workers exceeding 1 mSv (RN1) decreased from 19.1% in 2009 to 6.5% in 2019, while the proportion exceeding 10 mSv (RN10) remained very low, dropping from 1.0% in 2009 to 0.1% in 2019. The Mann-Kendall test revealed a statistically significant reduction in average annual doses (P < 0.01), with higher doses observed in specific categories, such as nuclear medicine, interventional radiology, industrial radiography, and well logging. These results suggest a consistent decrease in radiation exposure, stabilizing at a low level, reflecting the effectiveness of radiation protection measures during this period. However, the findings highlight the need for ongoing monitoring, especially for workers in high-exposure occupations, to ensure effective radiation protection.

  PublisherDOI

• When Graph Contrastive Learning Backfires: Spectral Vulnerability and Defense in Recommendation

  ArXiv.org · 2025-07-10

  preprintOpen accessSenior author

  Graph Contrastive Learning (GCL) has demonstrated substantial promise in enhancing the robustness and generalization of recommender systems, particularly by enabling models to leverage large-scale unlabeled data for improved representation learning. However, in this paper, we reveal an unexpected vulnerability: the integration of GCL inadvertently increases the susceptibility of a recommender to targeted promotion attacks. Through both theoretical investigation and empirical validation, we identify the root cause as the spectral smoothing effect induced by contrastive optimization, which disperses item embeddings across the representation space and unintentionally enhances the exposure of target items. Building on this insight, we introduce CLeaR, a bi-level optimization attack method that deliberately amplifies spectral smoothness, enabling a systematic investigation of the susceptibility of GCL-based recommendation models to targeted promotion attacks. Our findings highlight the urgent need for robust countermeasures; in response, we further propose SIM, a spectral irregularity mitigation framework designed to accurately detect and suppress targeted items without compromising model performance. Extensive experiments on multiple benchmark datasets demonstrate that, compared to existing targeted promotion attacks, GCL-based recommendation models exhibit greater susceptibility when evaluated with CLeaR, while SIM effectively mitigates these vulnerabilities.

  PublisherOA PDFDOI

• FAIR-SE: Framework for Analyzing Information Disparities in Search Engines with Diverse LLM-Generated Personas

  2025-11-08

  articleOpen access

  Search engine personalization, while enhancing user satisfaction, can lead to information disparities. Previous studies on this topic face limitations, such as the absence of context-aware data collection, superficial URL-level analysis, and human-dependent annotations. We propose FAIR-SE, a Framework for Analyzing Information dispaRities in Search Engines that addresses these challenges through AWS Lambda-based concurrent data collection and LLM-generated persona-based content analysis. We collected search results across four user contexts (Search History, Geo-location, Language Preference, and Access Environment) and analyzed them through four analytical perspectives (Political Leaning, Topic-specific Stance, Subjectivity, and Bias). Experiments conducted on two globally prominent search engines across nine controversial topics demonstrate the efficacy of FAIR-SE regarding benchmark accuracy, persona consistency, and ability to reflect real-world discourse patterns across diverse topics. Our statistical analysis identifies distinct search engine characteristics and demonstrates significant information disparities in our case studies examining regional disparities in search results. Our code and datasets are publicly available at: https://github.com/bigbases/FAIR-SE.

  PublisherDOI

• From intention to implementation: automating biomedical research via LLMs

  Science China Information Sciences · 2025-06-23 · 18 citations

  preprintOpen access
  PublisherOA PDFDOI

Recent grants

• EAGER: SaTC-EDU: Privacy Enhancing Techniques and Innovations for AI-Cybersecurity Cross Training

  NSF · $300k · 2020–2023

• NetSE: Medium: Privacy-Preserving Information Network and Services for Healthcare Applications

  NSF · $1.1M · 2009–2014

• TWC: Medium: Privacy Preserving Computation in Big Data Clouds

  NSF · $1.2M · 2016–2022

• A Peer to Peer Approach to Large Scale Information Monitoring

  NSF · $250k · 2003–2008

• CT-ISG: Protecting Location Privacy in Location-Aware Computing: Architectures and Algorithms

  NSF · $350k · 2006–2010

Frequent coauthors

• Arun Iyengar

  442 shared

• Karl Aberer

  408 shared

• Zhaohui Wu

  Wuhan Ship Development & Design Institute

  408 shared

• Weisong Shi

  University of Delaware

  406 shared

• Elisa Bertino

  Purdue University System

  405 shared

• M. Brian Blake

  Micropharma (Canada)

  405 shared

• Dimitrios Gerogakopolous

  Georgia Institute of Technology

  405 shared

• James Joshi

  University of Pittsburgh

  324 shared