Digital Commons Network^™

When Saying "No" Is Not Enough: Cognitive-Action Decoupling And The Illusion Of Safety In Llm Agents, Shasha Yu, Fiona Carroll, Barry L. Bentley

School of Professional Studies

Current safety evaluations of large language models (LLMs) predominantly rely on textual compliance, implicitly assuming that refusal-style responses correspond to safe behavior. This assumption becomes fragile when LLMs are embedded in agentic systems with the ability to execute state-changing actions. In this paper, we present an empirical critique of text-centric safety evaluation through an action-aware study of LLM agents under controlled conditions. Across multiple state-of-the-art models, we observe a recurring cognitive–action decoupling: agents generate policy-aligned refusal language while still producing unsafe tool-mediated action proposals. This produces an illusion of safety, where conversational audits indicate compliance even as operational risk persists. …

Systematic Approaches To Characterizing Vulnerabilities And Enhancing Robustness Of Text And Vision-Language Models, Poojitha Thota

Computer Science and Engineering Dissertations

The proliferation of artificial intelligence (AI) across critical domains, including news summarization, privacy-policy analysis, and medical decision support, has raised growing concerns about the security and robustness of these systems against adversarial manipulation. This dissertation investigates adversarial robustness in generative AI by addressing three key research goals: (1) characterizing adversarial vulnerabilities across generative models, (2) developing systematic defenses to improve the robustness of generative models, and (3) designing deployment-time safeguards for securing LLM interactions.

Towards the first goal, we characterize adversarial vulnerabilities across text-based and multimodal systems. In abstractive text summarization, we show that inference-time perturbations can exploit lead bias …

Beyond Fixed Thresholds: Per-Label Calibration For Fine-Grained Emotion Detection On The Goemotions Dataset, Sai Puneet Naga Venkata Subramanyam Patchipulusu

Selected Full-Text Master Theses 2021-

This study investigates the effectiveness of five community fine-tuned transformer models for fine-grained emotion detection on the GoEmotions dataset: SamLowe/roberta- base-go_emotions (RoBERTa-base), cirimus/modernbert-base-go-emotions (ModernBERT), mrm8488/deberta-v3-base-goemotions (DeBERTa-v3-base), bhadresh-savani/bert-base-go- emotion (BERT-base-cased),and tasinhoque/distilbert-go-emotions (DistilBERT) . While the original GoEmotions research by Demszky et al. (2020) established a BERT-base baseline with a macro-F1 of 0.46, this thesis extends that work through independent empirical evaluation of five derivative models, systematic per-label threshold optimization, and comparative analysis of architectural trade-offs across the full transformer model family. Using the GoEmotions simplified test split (5,427 examples across 28 categories), all five models were evaluated at a fixed 0.5 …

Previsit Ai: A Retrieval-Augmented Generation For Patient Readiness In Clinical Encounters, Rolande Umuhoza

All Graduate Theses, Dissertations, and Other Capstone Projects

With healthcare systems under growing pressure from rising patient volumes and shrinking consultation windows, improving how patients communicate with physicians has become essential to delivering quality care. Yet patients routinely arrive at appointments unable to clearly describe their symptoms, recall their medical history, or articulate concerns, contributing to miscommunication, diagnostic inefficiency, and pre-visit anxiety. This study introduces PreVisit AI, a conversational system designed to address this gap through structured, knowledge-based patient preparation. The system is built on a Retrieval-Augmented Generation (RAG) architecture combining HuggingFace sentence embeddings (all-MiniLM-L6-v2), a Chroma vector store, and Google’s Gemini language model over a curated seven-document …

Toward Interpretable Multi-Omics Multimodal Biomedical Artificial Intelligence, Yanjun Lyu

Computer Science and Engineering Dissertations

The complexity of human disease arises from biological processes that unfold across multiple scales, from molecular variation through cellular function, tissue organisation, brain phenotypes, each of which is associated with distinct measurement modalities, regularities, and characteristic. Contemporary biomedical artificial intelligence has brought the opportunity to reveal the complexity with in; however, its methodological default, in which models are trained on most readily available modality, does not adequately engage with the multi-scale connected structure by which biological meaning is constituted. The research area of multi-omics and multi-modal AI for biomedicine remains at an early exploratory stage, and the work presented in …

Examining Inclusive Computing Education For Blind Students In India, Akshay Kolgar Nayak, Yash Prakash, Md Javedul Ferdous, Sampath Jayarathna, Hae-Na Lee, Vikas Ashok

Computer Science Faculty Publications

The growing demand for computer professionals, driven by the expanding Information Technology industry, has led to numerous inclusive computing education efforts. These efforts have even included blind or visually-impaired (BVI) students, who are being increasingly encouraged to pursue education and a career in computing, despite the visually-oriented nature of the discipline. Extant literature has predominantly focused on identifying and addressing the accessibility barriers faced by BVI students to promote more inclusive learning environments. While few studies have also investigated the accessibility of computing education from the perspectives of BVI learners and instructors, these have been primarily situated in the Global …

Exploring Large Language Models For Trustworthy Use: Insights From Research And Development, Sandeep Kalari, Sahithi Padidela, Vikas Ashok, Ravi Mukkamala

Computer Science Faculty Publications

Large Language Models (LLMs) are increasingly being adopted in a wide variety of domains, including sensitive domains such as healthcare and finance. However, persistent challenges such as unreliable data sources, privacy breaches, and hallucinated output continue to hinder their usage. We have experimented with several strategies to address these challenges. First, we developed BlockQwen, a blockchain-augmented framework that integrates decentralized trust validation, role-specific access control, and verifiable audit trails into the Qwen 2.5 LLM workflow. Second, we developed PrivAware, a multilayered privacy-enforcement framework, using a fine-tuned Flan-T5 model with self-attention masking, to safeguard data while maintaining high utility. Both systems …

Attention-Based Geo–Textual Fusion Network For Disaster Risk Prediction, Mohammad Shafat Ahsan, Mst Sanjida Alam, Syed Sajjad Ahmed, Md Tanzimul Islam, Hashibul Ahsan Shoaib, M.F. Mridha, Md. Jakir Hossen

Student Publications [Scholarly]

Natural disasters pose recurring threats to human life and infrastructure, demanding intelligent systems that can process heterogeneous data streams and provide actionable insights in real time. Existing approaches often treat textual signals from social media and emergency communications separately from spatial hazard attributes, limiting their effectiveness in capturing the full complexity of evolving crises. This paper proposes an AI-driven geo–textual intelligence framework that integrates disaster-related text with GIS-based hazard features for real-time risk prediction and evacuation planning. The framework employs contextual text encoders and a neural GIS encoder, fused through an attention mechanism that dynamically weights cross-modal signals. Experiments on …

Llm-Assisted Legal Propositions Identification From Party Arguments In The U.S. Supreme Court Briefs, Heng Zheng, Alex Zhang

Faculty Scholarship

Merits briefs are central to U.S. litigation, serving as the primary means for parties to present arguments and persuade judges. Legal propositions in these merits briefs are the atomic units of arguments, whose relationships evolve throughout litigation and inform court decisions and precedent. Large language models (LLMs) have been applied to legal document review, but there is limited evidence on their ability to identify legal propositions in merits briefs. Given the labor-intensive nature of the task, we evaluate a human-AI collaborative approach to identifying legal propositions in the U.S. Supreme Court merits briefs, in which legal annotators review and revise …

Predicting Next-Day Eur/Usd Direction Using News Sentiment And Technical Indicators With Finbert And Xgboost, Darshan Sanjaybhai Khunt

Selected Full-Text Master Theses 2021-

Forecasting exchange-rate movements is a challenging task because currency prices are influenced not only by macroeconomic and financial variables but also by market sentiment reflected in financial news. This thesis examines whether financial news headlines can be used to predict the next-day directional movement of the EUR/USD exchange rate by applying finance-specific natural language processing and machine learning techniques.

The study uses a dataset of approximately 466,000 finance-related English-language news headlines collected between 2021 and 2025, aligned with daily EUR/USD closing prices. After preprocessing and temporal alignment, the data are used to construct a binary classification task in which the …

Simulating Social Attitudes With Llms: Accuracy, Demographic Effects, And Refusal Behavior In The Sensitive Domain Of Suicide Prevention, Cristina J. Perez, Michael P. Vasquez Jr., Philippe J. Giabbanelli, Patrick Y. Wu

VMASC Publications

Large language models (LLMs) are increasingly used to simulate public opinion, yet their validity in sensitive policy domains remains underexplored. We evaluate whether LLMs can reproduce attitudes toward suicide prevention policies using 32 questions drawn from seven nationally representative U.S. surveys (2023-2025). We systematically vary demographic conditioning (race/ethnicity, gender, age, education, income, party), prompt framing (direct elicitation, respondent embodiment, specialist embodiment), and model architecture (GPT-5 Nano, DeepSeek V3.2, Meta Llama 3.1 8B, Mistral Small 24B). Across 811,560 prompts, the mean absolute error—the average gap between predicted and human response distributions—is 23 percentage points. We also find that LLM responses to …

Advancing Hate Speech Detection: Binary And Multiclass Approaches From Traditional Methods To Parameter-Efficient And Ontology-Guided Language Models, Mahmoud Abusaqer

Graduate Theses/Dissertations

The widespread proliferation of hate speech on social media platforms poses significant challenges for content moderation and user safety, requiring automated systems that are simultaneously accurate, efficient, and capable of fine-grained distinctions. This thesis investigates hate speech detection through five published manuscripts organized into two complementary threads: binary detection (hateful vs. non-hateful) and multiclass detection across demographic targeting categories. The binary thread progresses from a broad 38-model baseline spanning traditional machine learning, deep learning, and transformer architectures (where RoBERTa reaches 91.48% accuracy and CatBoost remains competitive at 88.60%) to parameter-efficient adaptation, in which Low-Rank Adaptation (LoRA) of large language models …

News Media Sentiment Toward Chinese Ai: A Comparative Analysis With Belt And Road Initiative Involvement And Public Opinion On China, Asya Vaisberg

Pomona Senior Theses

This study evaluates different countries' news media’s sentiment towards Chinese AI, between May 2023 and May 2024, by using Microsoft Azure NLP Sentiment Analysis. The results are then compared with the country’s public opinion on China and its involvement in the Belt and Road Initiative (BRI). For this study 12 countries have been selected which are USA, Australia, Pakistan, Peru, Russia, Romania, Italy, Greece, Portugal, Philippines, Brazil, and Egypt. For each one, GNews Application Programming Interface (API), which has access to more than 60,000 global news sources, was used to aggregate relevant news articles based on queried keywords. The collected …

An Llm-Driven System For Doctor-Patient Simulation, Akilan Amithasagaran

Computer Science Theses

Effective physician-patient communication is fundamental to clinical competence, yet traditional simulation-based training methods using standardized patients and high-fidelity manikins are costly, resource-intensive, and difficult to scale. This dissertation presents CLiVR (Conversational Learning system in Virtual Reality), an LLM-driven system that integrates large language models and 3D avatars to simulate doctor-patient interactions for medical communication training.

CLiVR addresses three key limitations in existing virtual reality medical training platforms. First, the system operates on standalone Meta Quest 3 hardware with realistic 3D patient avatars featuring synchronized lip movements and speech-based interaction. Second, CLiVR grounds LLM responses using a curated syndrome-symptom database, constraining …

Bail Reform, Large Language Model Risk And Reasoning, William Wyatt

CGU Theses & Dissertations

This dissertation contains three studies. Each asks how rules or language change the choices people and machines make when outcomes are uncertain. The first study, written with Kiran John, evaluates California’s 2020 cashless bail reform. We use propensity score matching on arrestee records from the windows before and after implementation, and we test whether the shift away from cash bail produced any effect on subsequent offending. It did not. Matched comparisons yield small, statistically insignificant differences across every window we examined. That null result cuts against both sides of the public argument. The reform did not drive a spike in …

Learning Design To Advance Human-Ai Collaboration In K-12 Education, Wing Sha Chan, Jinhee Kim, Seongryeong Yu, Rita Kay Detrick

STEMPS Faculty Publications

This chapter explores key components for designing effective Human-AI Collaboration (HAC) in K–12 education, addressing the current lack of theoretical and conceptual frameworks for structuring and implementing HAC in teaching and learning. It examines four essential areas: curriculum design, student and teacher–AI interaction, learning environments, and the evolution of HAC over time. The chapter introduces the concept of HAC in K–12 contexts, highlighting how humans and AI can leverage each other's strengths through co-evolutionary processes that foster mutual learning and collaboration. It reviews current HAC practices in schools and discusses their contributions to both teaching and learning. Finally, it presents …

Do Developers Read Type Information? An Eye-Tracking Study On Typescript, Samuel W. Flint, Robert Dyer, Bonita Sharif

Research & Publications

Statically-annotated types have been shown to aid developers in a number of programming tasks, and this benefit holds true even when static type checking is not used. It is hypothesized that this is because developers use type annotations as in-code documentation. In this study, we aim to provide evidence that developers use type annotations as in-code documentation. Understanding this hypothesized use will help to understand how, and in what contexts, developers use type information; additionally, it may help to design better development tools and inform educational decisions. To provide this evidence, we conduct an eye tracking study with 26 undergraduate …

Adaptive Boundary-Aware Fact-Checker Placement For Misinformation Suppression In Social Networks, Mostafa Taghizade Firouzjaee, Ghazal Naderi, Ross Gore, Neda Moghim

School of Cybersecurity Faculty Publications

The spread of fake news on online social networks is driven by imitation-based user behavior and network topology, often leading to persistent misinformation clusters and echo chambers. In this study, we develop a spatial evolutionary game-theoretic framework in which agents update their latent opinions through payoff-biased imitation, while external fact-checkers act as non-imitative intervention nodes. Building on this formulation, we propose an adaptive, boundary-aware intervention mechanism that dynamically regulates both the density and spatial allocation of fact-checkers according to real-time system conditions. Competing information clusters are identified through local neighborhood composition, enabling boundary nodes, i.e., interfaces between fake-news and non-fake-news …

Thinkmatter: Panoramic-Aware Instructional Semantics For Monocular Vision-And-Language Navigation, Guangzhao Dai, Shuo Wang, Hao Zhao, Bin Zhu, Qianru Sun, Xiangbo Shu

Research Collection School Of Computing and Information Systems

Vision-and-Language Navigation in continuous environments (VLN-CE) requires an embodied robot to navigate the target destination following the natural language instruction. Most existing methods use panoramic RGB-D cameras for 360° observation of environments. However, these methods struggle in real-world applications because of the higher cost of panoramic RGB-D cameras. This paper studies a low-cost and practical VLN-CE setting, e.g., using monocular cameras of limited field of view, which means “Look Less” for visual observations and environment semantics. In this paper, we propose a ThinkMatter framework for monocular VLN-CE, where we motivate monocular robots to “Think More” by 1) generating novel views …

Memoryart: Enhancing Llms Via Multi-Memory Models With Adaptive Resonance Theory For Healthcare Agents, Renke Dai, Hebin Hu, Jiahui Zhang, Yilin Kang, Ah-Hwee Tan

Research Collection School Of Computing and Information Systems

Though promising in healthcare consultation applications, large language models (LLMs) face critical limitations in retaining and utilizing long-term memory across multiturn interactions. In particular, existing memory enhancing paradigms are constrained by limited context windows and embedding-based retrieval, often failing to maintain task relevance and still suffering from memory prototype collapse in multi-turn healthcare consultation. To address these challenges, we propose a cognitively-inspired memory framework named MemoryART, which is grounded in Adaptive Resonance Theory (ART)—a cognitive and learning theory of how humans and animals adapt to dynamic environments. MemoryART employs three memory modules—working memory, episodic memory, and semantic memory to support …

Clique Annealing: Semi-Supervised Community Detection Under Crystallization Kinetics, Ling Cheng, Jiashu Pu, Ruicheng Liang, Qian Shao, Hezhe Qiao, Feida Zhu

Research Collection School Of Computing and Information Systems

Semi-supervised community detection seeks to find a specified community type when only few communities are labeled. Existing "select-then-refine" pipelines often start from mis-aligned cores and rely on Reinforcement-Learning or Generative Adversarial Network, increasing computational cost and limiting scalability. We address these issues with a unified energy framework under crystallization kinetics that jointly models energy, structure, and growth. Based on this perspective, we propose CLique ANNealing (CLANN), which first employs Nucleus Proposer to select candidate clique as community core under four physics-inspired criteria. A learning-free Transitive Annealer then iteratively merges neighboring cliques and repositions the nucleus, enabling spontaneous, scalable community growth. …

Understanding Phishing Susceptibility Through Expert Consensus Using Digital Marketing Parallels And A Machine Learning-Based Implementation, Mansoor Ahmad

All Graduate Theses, Dissertations, and Other Capstone Projects

Phishing remains one of the most effective attack vectors for gaining unauthorized access to organizational systems, yet defenders often lack systematic methods to assess their exposure before an attack. This study develops a framework that uses machine learning to encode the collective expertise of cybersecurity practitioners into a portable phishing susceptibility assessment tool, with the goal to help security teams proactively identify patterns, prioritize awareness training, and strengthen detection controls. The study surveyed 27 practitioners with extensive experience in social engineering, red teaming, penetration testing, and threat analysis to identify which factors most influence phishing susceptibility. Practitioners provided quantitative ratings …

Ai Vs. Genai: Combating Llm Generated Prescription Fraud With Transformer Based Detection Models, Ankitha Vokkaleri Shankarappa

Selected Full-Text Master Theses 2021-

The rapid evolution of Large Language Models (LLMs) has introduced a sophisticated new vector for healthcare fraud: the generation of high-fidelity, synthetic medical prescriptions. Traditional fraud detection systems, which rely on rule-based engines and basic statistical anomalies, are increasingly ill-equipped to identify these AI-generated forgeries that mimic the structural and clinical logic of authentic records. This thesis presents a robust detection framework using Transformer-based architectures to distinguish between human-authored Medicare Part D prescriptions and fully synthetic records generated by GPT-4.

The research was conducted across two distinct phases: an initial pilot study using 4,000 samples and a rigorous validation stress …

A Comprehensive Survey On Facial Expression Generation: From Gans To Llm-Guided Multimodal Models, Murad Hasan, Rabab Abdelfattah, Kareem Abdelfatah, Mostafa Fouda, Ahmed Sherif

Faculty Publications

Facial expression generation (FEG) has emerged as a vital area in human–computer interaction, virtual avatars, and affective computing, aiming to synthesize natural and expressive facial behaviors across diverse interaction contexts. This survey presents a comprehensive analysis of recent advances in FEG, organized into six key paradigms: speech-driven expression generation, facial reaction generation, face video generation, facial animation, avatar-based generation, and text-driven expression generation. We review a wide range of model architectures, including VQ-VAEs, Generative Adversarial Networks (GANs), 3D Morphable Models (3DMMs), Transformers, and diffusion-based approaches, and compare their performance using commonly adopted evaluation metrics such as Fréchet Distance (FD), Peak …

Towards Breach Hypothesis Based Predictive Autonomous Cyber Defense Ecosystem: Proactive Prevention Of Imminent Threats And Productivity Losses, Yogesh Chavarkar

Master's Theses and Doctoral Dissertations

As cyberattack tools and techniques get sophisticated and persistent, reactive cybersecurity has been unable to effectively prevent breaches and compromises. Organizations and institutions with large complex environments have wide vulnerable exposure with higher chances of a cyberattack. This also increases overall security and financial risk. Possibility of repetitive attacks from cyber threats increases, too, despite the use of standard defense measures. Security tools and system vulnerabilities often need manual patching and updates to keep the environment secure and functioning effectively. Productivity suffers from manual trade-offs in keeping the systems secure from adverse impact. Persistent high-severity cyberattacks, despite continued reactive defensive …

Emotional Patterns In Conversational Social Media Using Graph-Based Context, Sai Sanjay Yerunkar

Master's Projects

Most social media datasets for emotion detection have not been constructed to account for conversational structure‚ so we investigate whether it carries signal for emotion prediction. Using Sentiment140 and GoEmotions Reddit threads‚ we construct their thread-based conversation graphs and compute the aggregated features of neighbors as well as the transformer and TF-IDF representations of comments. Connected comments are 4.5× more similar in emotions than expected by chance. Pairwise emotional similarity decays exponentially with geodesic distance (e.g.‚ after 2 hops). Pairs separated by a 30s timestamp difference have the highest emotional similarity. These results suggest that emotion is structured locally and …

Handwriting Recognition In Vr, Dominique Mosley

EWU Masters Thesis Collection

Virtual Reality (VR) is slowly becoming more popular for more than just entertainment. VR can be found in educational, office, and even healthcare settings to help discover more intuitive ways to teach, collaborate, and treat patients. Outside of the virtual world, these environments typically rely on writing for communicating or note-taking. Currently, VR input forces users to rely on clunky on-screen keyboards which disrupts the user’s immersion and breaks the flow of natural interaction. This thesis explores the potential of VR as a learning platform by combining it with artificial intelligence (AI). It aims to develop a VR-enhanced handwriting practicing …

The Case For Friction In Ai-Mediated Information Seeking And Learning, Susan G. Archambault, Priya Kizhakkethil

Librarian Publications & Presentations

Introduction. This paper challenges the assumption that frictionless AI-mediated information seeking represents progress. We argue that AI systems eliminating productive friction, such as uncertainty, exploration, and reflective processes, undermine intellectual virtues essential for critical thinking in an AI-saturated information environment.

Method. This conceptual article employs a theory synthesis approach, drawing on library information science (LIS) theories of information behaviour, experience, and literacy, virtue epistemology, and human–computer interaction (HCI) friction design literature.

Analysis. We map intellectual virtues such as curiosity, thoroughness, and intellectual humility onto the Association of College and Research Libraries (ACRL) Framework for Information Literacy for Higher Education, building …

Empirically Evaluating The Accessibility Of A Pon-Enable Feature Diagrams Notation By The Red-Green Colorblind Community, Mohamed El-Attar, Sarah Kohail, Rima Grati

All Works

In 2016, an enhanced version of a feature diagram notation developed using the Physics of Notations (PoN) framework was introduced. Empirical evidence demonstrated that this revised notation was more cognitively effective than the original. However, the new notation relies on color, specifically red, which poses accessibility challenges for individuals with red–green color vision deficiency, as they cannot perceive the notation as originally intended. Consequently, the cognitive effectiveness of a red–green–deficient (RGD) version of the new notation relative to the original notation remained unknown. Although the PoN framework specifies several principles that may be satisfied with or without the use of …

Mapping Multiclass-Targeted Hate Speech In Online Discourse: An Open Dataset, Sanaa Kaddoura, Sumaia Al-Kohlani

All Works

Online social networks have become central spaces for public discourse, where hostile and discriminatory language toward social groups can cause psychological and social consequences for marginalized communities. Although multiple public hate speech datasets are available, many rely on binary categorization practices that obscure linguistic, cultural, and contextual variation across targeted groups. As a result, minority and less visible forms of hate speech remain insufficiently documented and analyzed. This discussion paper examines methodological limitations in existing hate speech annotation schemes and presents a re-annotation framework applied to the HatEval2019 dataset. The proposed framework introduces target-specific multiclass labels that distinguish subcategories of …