- Gradstudents - mailman.cs.vt.edu

Waris Gill's Final PhD Defense, 3:45 PM, Friday, December 5
by Muhammad Ali Gulzar 03 Dec '25

03 Dec '25

Hi everyone, I am pleased to invite you to Waris Gill’s final PhD defense at 3:45 PM on Friday, December 5 in 4307 Gilbert Place. The Zoom link is: https://virginiatech.zoom.us/j/88487256025 Title: Interpretability and Debugging for Distributed Privacy Preserving Machine Learning Abstract: Machine Learning (ML) is transforming numerous fields, with Large Language Models (LLMs) driving applications ranging from personalized healthcare to intelligent chatbots. However, training data required by these models is not always shareable or centralized due to privacy concerns (e.g., GDPR, HIPAA) and requires significant computational resources. To address this, Federated Learning (FL) has emerged as a distributed privacy-preserving ML paradigm. FL trains a global model without centralizing client data by performing local training on client devices and aggregating updates on a central server, as seen in applications like Siri, Alexa, and healthcare. While FL offers privacy benefits, it sacrifices interpretability. Malicious or faulty client models can degrade global model quality, and attributing predictions to specific clients is challenging. This thesis investigates how we can design automated debugging and interpretability methods for federated learning that effectively localize faults and attribute global model predictions without degrading performance or violating core FL's privacy principles. This thesis, first, redefines differential testing and neuron-level model behaviors to exclude or limit faulty clients due to unintentional issues (e.g., labeling errors) or intentional attacks (e.g., poisoning attacks), improving FL training reliability. Second, the thesis introduces the concept of fine-grained neuron provenance, which enables the identification of each client's contribution, revealing which participants are primarily responsible for particular predictions of the global model. Third, the research extends FL interpretability to Large Language Models (LLMs) with a novel token-level provenance attribution method that leverages per-token activation-gradient analysis to quantify each client's influence on individual tokens while preserving privacy constraints. This approach establishes a foundation for interpretable and trustworthy federated LLM systems. Collectively, all three investigations culminate in practical tools that integrate seamlessly with existing distributed ML workflows. These tools enable real-time debugging, transparent model insights for both classification and generative models, and reduced LLM inference costs. Best, Gulzar — Muhammad Ali Gulzar Assistant Professor Virginia Tech | Department of Computer Science 220 Gilbert Street, Room 4106 Blacksburg, VA 24060 (540) 231-0851 | gulzar(a)cs.vt.edu | https://people.cs.vt.edu/~gulzar

1 0

CS Student Experiences Survey
by Moussad, Bernard 03 Dec '25

03 Dec '25

Dear Computer Science Graduate Student, The Virginia Tech Department of Computer Science (CS@VT) is continuously working to improve the culture and quality of graduate programs. In order to do so, we rely on our students’ honest feedback about what they experience each semester. What we learn from this feedback allows us to recognize and build upon our areas of strength and to allocate time and resources to remediate those areas that require attention. This data will be collected anonymously and it will only be used for the purpose of strategic improvement. A report of the aggregated data will be shared with department administrators to be used when determining strategic priorities for the program and to facilitate transparency. If you have any questions about the survey, please contact the CS Grad Council: bernardm(a)vt.edu https://forms.gle/BiieFZCSPWEcCTWx5 Sincerely, Bernard Moussad CSGC President PhD student in CS@VT Phone: (256) 668-7571 bernym12.github.io<http://bernym12.github.io>

1 0

Invitation to Badhan Das' final defense, 2:00 PM, December 5
by Lenwood S. Heath 02 Dec '25

02 Dec '25

I am please to invite you to attend the final PhD defense of Badhan Das at 2:00 PM EDT on Friday, December 5, in Torgersen 3310. The title of her dissertation is Graph-Based Computational Approaches for Modeling Viral Evolution and the topic is computational epidemiology combining viral evolution with graph theory. The Zoom link is https://virginiatech.zoom.us/my/lennyheath The abstract is below my signature. See you there! -- Lenwood S. Heath Professor 2160J Torgersen Hall https://people.cs.vt.edu/~heath/ Department of Computer Science Cell phone: (540) 392-6672 Virginia Tech FAX: (540) 231-6075 Blacksburg, VA 24061-0106 Email: heath(a)vt.edu -------------------------------------------------------------------------- Message and signature (c)2025 Lenwood S. Heath All rights reserved -------------------------------------------------------------------------- Graph-Based Computational Approaches for Modeling Viral Evolution Badhan Das Abstract Modeling viral evolution is essential for understanding how pathogens adapt, spread, and give rise to new variants of concern, yet it remains challenging due to high mutation rates, minimal sequence divergence, and the enormous scale of genomic data. Classical network epidemic models typically treat the pathogen as genetically static during transmission, while traditional phylogenetic trees impose strictly bifurcating relationships that struggle to capture virus-specific evolutionary phenomena such as recurrent and back mutations, convergent evolution, and recombination. In contrast, quasispecies theory describes viral populations as clouds of closely related mutants that occupy a high-dimensional sequence space, where connectivity is naturally represented as a graph rather than a tree. This graph-based view captures richer mutational relationships and provides a richer representation of the evolutionary processes shaping RNA viruses such as SARS-CoV-2. In this dissertation, I develop a sequence of graph-centered frameworks that integrate viral fitness, evolutionary distance, and mutational dynamics to model viral evolution from mechanistic, data-driven, and predictive perspectives. ViraFit introduces a proof-of-concept model that couples epidemiological spread on a contact network with evolutionary dynamics on a fitness landscape, incorporating fitness landscapes, viral quasispecies, and genome structure to simulate the diffusion of a mutating virus across hosts. By tuning genome length, network size, mutation rate, infection probability, and infection time, ViraFit shows that while the fitness model shapes evolutionary trajectories, the underlying network structure also modulates landscape ruggedness and the adaptive walk toward the fittest strain. Building on real genomic data, the Variant Evolution Graph (VEG) provides a scalable graph-based representation of SARS-CoV-2 evolution based on mutational distances. Unlike traditional phylogenetic trees, VEG allows multiple ancestors per variant, captures diverse evolutionary pathways, and reveals recombination, homoplasies, and intra-host diversity through strongly connected subgraphs. A derived Disease Transmission Network further supports inference of transmission links and potential superspreaders. Applying VEG to genomes from Somalia, Bhutan, Hungary, Iran, and Nepal, and comparing them with Maximum Likelihood phylogenies, I demonstrate that VEG offers a computationally efficient alternative for large-scale viral analysis. Extending these ideas, the Mutation Learning Graph (MLG) organizes viral genomes into a directed, mutation-aware structure defined by cumulative mutation profiles relative to a reference. Using a novel alignment-aware Ancestor-Joining algorithm, MLG infers intermediate ancestors to restore mutational continuity and reconstruct cohesive quasispecies-like clouds. Across ten geographically diverse SARS-CoV-2 cohorts, Egypt, Iran, Nigeria, Bangladesh, Queensland, China, Estonia, Wyoming, Chile, and South Africa, MLG supports node-level lineage classification and edge-level link prediction. Node classification serves as a diagnostic for feature expressiveness, while link prediction confirms that MLG encodes biologically meaningful mutational transitions, with GGNN exhibiting the strongest overall performance. Together, ViraFit, VEG, and MLG form a unified methodological progression linking mechanistic simulation, data-driven evolutionary reconstruction, and predictive graph learning, providing a scalable, interpretable, and mutation-centric view of viral evolution that complements traditional phylogenetic approaches and enables future variant forecasting.

1 0

Final PhD defense of Satvik Chekuri at 11am, Tuesday, December 16 in 2030 Torgersen Hall and in Zoom
by Fox, Edward 02 Dec '25

02 Dec '25

Hi! I am pleased to invite you to Satvik Chekuri's final PhD defense at 11am on Tuesday, December 16, in 2030 Torgersen Hall. The Zoom link is: https://virginiatech.zoom.us/my/ed.fox Please note that topics covered include AI, NLP, LLMs, Question Answering, Knowledge Graphs, Information Retrieval (aka search engines), Digital Libraries, User Studies, ETDs, Topic Modeling, Question Generation, Object Detection, Ontologies, Automatic Classification, and Scholarly Publishing. Title: Scholarly Information System for Long Documents and their Elements: Structured Representation and Exploration using Knowledge Graphs and LLM Abstract: Electronic Theses and Dissertations (ETDs) contain some of the deepest and most detailed accounts of academic work, yet the systems used to access them remain limited. Traditional information retrieval emphasizes document-level search and provides only coarse entry points into long-form scholarly writing. Large language models (LLMs), although capable of fluent explanation, lack grounding in the structure of ETDs and cannot reliably reason over their internal elements. This dissertation addresses these challenges by developing a scholarly information system that integrates fine-grained retrieval, structured knowledge graph representations, and grounded question answering to support richer access to ETD knowledge. Instead of treating an ETD as a single monolithic text, the proposed ETD-IR system (Information Retrieval System for ETDs) extracts and indexes elements like figure captions, topic labels, and abstract sentences, allowing users to retrieve precise evidence. We introduce a formal ontology to capture semantic relationships in ETDs, leading to the creation of the first ETD-specific Knowledge Graph (ETD-KG). To enhance scholarly discovery and question-answering, this dissertation proposes ETD-OHKGR, an ontology-guided hybrid KG retrieval framework. ETD-IR retrieves relevant documents, ETD-IR + LLM synthesize results with a LLM, and ETD-OHKGR adds graph-based reasoning to generate answers based on textual and structural evidence. This research contributes an ETD Abstract Sentence Classification dataset (ETD-ASC), consisting of human-labeled and LLM-assisted silver annotations. Fine-tuned models trained on this dataset highlight that ETD abstracts exhibit rhetorical patterns distinct from those found in scientific papers, reinforcing the need for ETD-specific classification models. These class labels are used to improve knowledge extraction and structural coherence within ETD-KG. A comprehensive user study involving 47 participants across four engineering departments has compared the three methods on informativeness, accuracy, readability, and clarity metrics. The ETD-OHKGR method has achieved 40% higher overall user satisfaction than ETD-IR (p < 0.001). It also delivered major improvements in presentation quality, with a 53% gain in readability and a 48% gain in clarity. ETD-OHKGR significantly outperformed both baselines on the two metrics that users consistently valued the most—informativeness and clarity. These improvements held for 87% of participants, demonstrating robust performance. To analyze the relationship between the responses from different methods, paragraph-level natural language inference (NLI) was used, treating ETD-IR+LLM as the premise and ETD-OHKGR as the hypothesis. ETD-OHKGR’s answers mainly fell into the neutral class (50–100%), suggesting it adds complementary content without contradictions, aligning with its goal of generating structured, elaborative answers based on ETD-KG. Semantic similarity analysis showed moderate divergence (mean cosine similarity = 0.420), emphasizing the role of structured retrieval with diverse writing styles. Strong correlations (r > 0.70) among metrics confirmed that clarity and readability predict user satisfaction. Taken together, these findings demonstrate that fine-grained retrieval, ontological structure, and knowledge graph reasoning enhance access to scholarly content in ETDs. The proposed system offers capabilities that neither traditional IR nor LLMs alone can achieve, making long-form academic documents more discoverable and useful for researchers. All are welcome! PS Special thanks go to Satvik's doctoral committee and to collaborators in University Libraries. -- Edward A. Fox, Ph.D., Emeritus Professor of Computer Science Virginia Tech | CS (0106) 3160 Torgersen Hall (CS dept. office) Blacksburg, VA 24061 (540) 231-5113 | fox(a)vt.edu | http://fox.cs.vt.edu

1 0

Tenure Track Position at RU (fwd)
by Lenwood S. Heath 02 Dec '25

02 Dec '25

Hello everyone, One of our graduates, Nathaniel Lahn, is a professor of computer science at Radford University. He wanted me to tell you about two tenure-track positions open in his department. See the email below and follow the link. -- Lenwood S. Heath Professor 2160J Torgersen Hall https://people.cs.vt.edu/~heath/ Department of Computer Science Cell phone: (540) 392-6672 Virginia Tech FAX: (540) 231-6075 Blacksburg, VA 24061-0106 Email: heath(a)vt.edu -------------------------------------------------------------------------- Message and signature (c)2025 Lenwood S. Heath All rights reserved -------------------------------------------------------------------------- ---------- Forwarded message ---------- Date: Mon, 1 Dec 2025 21:21:35 +0000 From: "Lahn, Nathaniel" <nlahn(a)radford.edu> To: "heath(a)vt.edu" <heath(a)vt.edu> Subject: Tenure Track Position at RU Resent-Date: Mon, 1 Dec 2025 16:21:48 -0500 (EST) Resent-From: <heath(a)vt.edu> Lenny, I hope you are doing well! We are looking for a couple of tenure track positions in our department at Radford University for Computer Science / Cybersecurity. I would appreciate it if you could share this job posting with anyone you think might be interested. https://jobs.radford.edu/jobs/assistant-professor-computer-science-and-cybe… ~Nathaniel

1 0

Adithya Harish Final MS Defense
by Eli Tilevich 01 Dec '25

01 Dec '25

Dear all, You are all cordially invited to attend Adithya Harish final MS defense. Title: Persona Reinforcement for Secure Programming AI Tutors: Adaptive Assistance in Action Date: December 4th Time: 12 pm Zoom link: https://virginiatech.zoom.us/j/6161682089 Committee members: Eli Tilevich (Chair), Chris Brown, Na Meng Abstract: Software powers nearly every aspect of modern life, but software insecurity can hurt both people and organizations. Learning how to write secure code is therefore essential, yet teaching this skill effectively remains a challenge. With the rise of generative AI tools such as ChatGPT, many students rely on these systems to receive direct solutions, thus undermining the learning process. This thesis introduces Secure Programming with Adaptive Reasoning Companion or SPARC for short. This AI-powered tutor helps students learn secure programming by guiding them step by step instead of providing direct answers. SPARC encourages reasoning and problem solving by offering hints that adapt to each student's skill level. The system is integrated with SecureCoder, a drill-and-practice platform for secure coding, through which students can apply these adaptive hints when solving problems. In a pilot study involving 20 participants, students using SPARC demonstrated greater success and engagement compared to those using a GPT model. Most participants described SPARC's hints as clear, helpful, and adaptive to their needs, noting that it deepened their understanding of secure coding. The detailed evaluation showed that SPARC effectively fulfills its goal as a secure programming tutor. By shifting the role of generative AI from one that undermines education to an active learning assistant, this research demonstrates how AI can effectively support rather than replace the educational process.

1 0

CS Graduate Program Newsletter: December Edition
by Dr. Eli Tilevich 01 Dec '25

01 Dec '25

Dear all, Please find attached the December edition of the Computer Science Graduate Program Newsletter, a monthly publication of the Graduate Program Office. Best regards, Dr. T. ---------------------------------------------------------------------- Eli Tilevich, Ph.D. Professor and Associate Dept. Head for Graduate Studies Gilbert Place #4204 Dept. of Computer Science Virginia Tech Zoom: https://virginiatech.zoom.us/my/tilevich Office hours: Wednesdays starting from 3:00pm (540) 231-3475 http://people.cs.vt.edu/~tilevich

1 0

REMINDER: THIS FRIDAY: Entrepreneurship and Computing Seminar Series: Tim Gorski: Nov 21st, 10:30am
by Kantarcioglu, Murat 20 Nov '25

20 Nov '25

Dear All, As part of the CS/Root, the hub for computer science entrepreneurship at Virginia Tech, Entrepreneurship and Computing Seminar Series, we are hosting a talk from Tim Gorski (an alumni of VT CS). We invite you to attend this talk if you are interested in startups and/or in learning how research can be successfully transitioned into practice. Please see the details below, along with the attached seminar announcement. Seminar Title: Entrepreneurship: Building a bootstrapped defense firm from 0 to 300 Date and Time: November 21, 2025, Friday, 10:30am. Location: 1100 Torgersen Speaker Bio: Tim Gorski is a founder, investor, and philanthropist with a proven track record of building and leading large scale software engineering and IT enterprises. He is the founder of Omynon, which he bootstrapped into a 300-person defense contracting firm serving mission-critical government and industry clients. A VT–trained software engineer and an entrepreneur at heart, he is recognized for guiding teams and organizations toward excellence through a blend of technical innovation, strategic insight, and purposeful leadership. Best Regards, Murat ==================================== Murat Kantarcioglu, Ph.D. Professor, and CCI Faculty Fellow Virginia Tech, Department of Computer Science www.kantarcioglu.net ====================================

1 0

GRA opening for Spring 2026: HCI/UX evaluation of OSINT software
by Luther, Kurt 19 Nov '25

19 Nov '25

Dr. Kurt Luther, Associate Professor of Computer Science at Virginia Tech, is seeking a Graduate Research Assistant (GRA) for Spring 2026, with possible continuation to Summer 2026. The GRA will support the work of the Crowd Intelligence Lab<https://crowd.cs.vt.edu/> and the OSINT Lab<https://hume.vt.edu/research/OSINTLab.html>. Duties Include: * Conduct formative and summative HCI/UX evaluations of a collaborative Open Source Intelligence (OSINT) software tool as part of a larger research team. * Assist in formulating and presenting design recommendations to the software development team, based on user feedback and findings from UX evaluations. Required Qualifications: * Current Virginia Tech graduate student planning full-time enrollment in Spring 2026. * Strong written and oral communication skills. * Due to funding requirements, you must be a U.S. citizen. Preferred Qualifications: * Earned degree in Computer Science, Business Information Technology, Cybersecurity, Industrial Engineering, Psychology, or related field. * Work experience or coursework in Human-Computer Interaction (HCI), Human Subjects Research, and/or UX research, design, and evaluation. * Work experience or coursework in Open Source Intelligence (OSINT), intelligence analysis, digital investigations, and/or cybersecurity. * Interest in military, law enforcement, or national security issues. To Apply: Please email Dr. Luther <kluther(a)vt.edu<mailto:kluther@vt.edu>> a brief cover letter outlining your relevant background, qualifications, and interest in the position, along with your resume / CV as a PDF attachment. Use the subject line “OSINT GRA application.” Applications are being accepted immediately until the position is filled. Thanks, Dr. Luther -- Kurt Luther, Ph.D. Associate Professor of Computer Science Virginia Tech https://crowd.cs.vt.edu/kurt-luther/

1 0

Seminar Talk: Friday, November 21, 2:30 - 3:45pm, Classroom Building, Room 260
by Zier, Debbie 19 Nov '25

19 Nov '25

Craig Zilles Professor and Severns Faculty Scholar Siebel School of Computer and Data Science University of Illinois at Urbana-Champaign Friday, November 21, 2025 2:30 - 3:45pm Classroom Building, Room 260 Seminar: Exams with More Learning and Less Stress with a Computer-based Testing Facility Abstract: Exams are an important tool for summative assessment, whose utility has only grown with the advent of large language models (LLMs) like ChatGPT, because they can be implemented in a trustworthy manner. But exams are generally not well liked by either students or faculty. Students find them stressful. For faculty (and their course staff), they represent a large administrative burden to write, proctor, and grade. This large burden means they are done infrequently in many classes, but this infrequent testing encourages cramming and leads to high test anxiety. In this talk, I'll share (1) research on the benefits of frequent testing and "second-chance testing" (optional exam re-takes) on increased student learning and decreased test anxiety, (2) research on patterns of cheating on unproctored online assessments, and (3) how we've reduced the instructor workload at Illinois to implement frequent testing through our Computer-based Testing Facility (CBTF). The CBTF is a collection of proctored computer labs that, in conjunction with the PrairieLearn open-source question-asking platform, enable our faculty to run sophisticated exams with almost no recurring effort even in the largest classrooms. For example, our CS 1 course for majors (run by a single faculty member) ran weekly exams for 1,150 students. Key enabling ideas for the CBTF include: (1) sophisticated auto-grading questions, (2) question generators, (3) asynchronous exams, and (4) dedicated testing space and proctors. The CBTF has been running for over 10 years and proctored over 100,000 exams last semester. Bio: Craig Zilles is a Professor and Severns Faculty Scholar in the Siebel School of Computer and Data Science at the University of Illinois at Urbana-Champaign. His current research focuses on applying computing and data analytics to education, including the development of the Computer-based Testing Facility (CBTF). Previously, his research focused on the interaction between compilers and computer architecture, and he developed the first algorithm that allowed rendering arbitrary three-dimensional polygonal shapes for haptic interfaces (force-feedback human-computer interfaces). He received the IEEE Education Society's 2010 Mac Van Valkenburg Early Career Teaching Award and an NSF CAREER award. At Illinois, he has received a wide range of teaching awards, including a 2018 Campus Award for Excellence in Undergraduate Teaching, a 2013 Illinois Student Senate Teaching Excellence Award, and the College of Engineering's Rose Award (2007) and Everitt Award (2008) for Teaching Excellence. He holds 5 patents and his research has been recognized by a best paper awards from ASPLOS in 2010 and 2013 and by selection for inclusion in the IEEE Micro Top Picks from the 2008 Computer Architecture Conferences. ________________________ Debbie Zier Administrative Assistant and Office Manager Department of Computer Science (0106) 3160E Torgersen Hall 620 Drillfield Drive Blacksburg, VA 24061 541-231-1369

1 0