Digital Commons Network^™

Cost-Aware Predictive Routing For Vision-Language Models, Sahil Sait Naveed

Master's Projects

Vision-language models (VLMs) are increasingly used for multimodal tasks, but their inference costs vary widely across model tiers. This work presents a predictive routing framework that assigns each query to the most cost-effective VLM by using multimodal embeddings, clustering, and per-cluster model error estimates. Evaluated on six commercial VLMs, the selected router reduces average cost by about 49 percent relative to the quality-first setting on both validation and test. This comes with accuracy losses of 0.4 percentage points on validation and 0.9 on held-out test data. It also outperforms K-NN and ZeroRouter routing baselines while providing lowercost operating points than …

Challenges And Applications Of Fine-Grained Temporal Action Understanding: Modeling Temporal Granularity And Data Efficiency, Halil I. Helvaci

Theses and Dissertations--Electrical and Computer Engineering

Fine-grained Temporal Action Segmentation (TAS) has become a cornerstone of video understanding, offering dense frame-level predictions essential for clinical assessment, surgical skill evaluation, and human-computer interaction. While TAS methods have delivered strong results on coarse-grained benchmarks, two fundamental challenges persist: (1) global attention mechanisms dilute boundary information critical for subsecond precision, a phenomenon we term the temporal granularity bottleneck, and (2) dense frame-level annotation remains prohibitively expensive, with most datasets requiring exhaustive labeling of lengthy untrimmed videos. These challenges are particularly pronounced in medical domains, where sub-second primitives define clinical outcomes while expert annotation remains scarce. In this dissertation, we …

Generative Ai Peer Tutoring To Support Peer-Reviewed Source Identification And Evaluation, Rae Mair, Michelle Kelley, Taylar Wenzel, Andrea C. Borowczak, Yuqing Li, Mike Borowczak

Faculty Scholarship and Creative Works

This chapter explores the design and evaluation of a generative artificial intelligence peer tutor prompt to support college students in identifying and evaluating peer-reviewed sources for academic research. Grounded in literature on peer tutoring, Socratic dialogue, and AI-supported learning, the authors describe an iterative prompt engineering process designed to transform large language models (LLMs) into Socratic-style peer tutors capable of scaffolding student reasoning without completing tasks for them. Five guiding criteria for an effective peer tutor shaped development and evaluation: cognitive congruence, step-by-step guidance, avoiding giving answers, adaptability to student level, metacognitive transparency, and following assignment directions. Across multiple human-centered …

Visual Analytics For Interpretable Quantum Computing, Shaolun Ruan

Dissertations and Theses Collection (Open Access)

Quantum computing has entered a stage of increasing practicality. Many quantum hardware vendors such as IBM, Rigetti, Honeywell, and IonQ now enable experiments on real devices in the Noisy Intermediate-Scale Quantum (NISQ) era. These platforms show computational advantages in domains such as optimization, machine learning, and materials science. However, they remain limited by hardware noise and the absence of human-interpretable information. Existing visual metaphors, such as the Bloch Sphere for single-qubit states or circuit schematics for algorithm design, struggle to convey multi-qubit entanglement or measurement probabilities in ways accessible to human reasoning. Likewise, the rise of variational quantum circuits and …

Constrained Reinforcement Learning: From Single-Agent Safety To Multi-Agent Coordination, Hao Jiang

Dissertations and Theses Collection (Open Access)

Real-world decision-making systems such as autonomous driving and largescale ride-pooling must operate under strict safety and resource constraints. Traditional Reinforcement Learning (RL) methods, while powerful in simulation, often fail to guarantee such constraints, limiting their real-world deployment. The fundamental challenge lies in integrating constraint satisfaction with long-term reward optimization, especially when outcomes are stochastic and interdependent across multiple agents.

This dissertation advances the field of Constrained Reinforcement Learning (CRL) from both single-agent safety and multi-agent coordination perspectives. In the single-agent setting, we introduce a Reward Penalty framework that augments the state space with cumulative cost and penalizes only trajectories that …

Improving Pid Control With Bayesian Optimization For Adversarially Robust Federated Learning, Adrian Pena, Sergei Chuprov, Raman Zatsarenko, Leon Reznik

Computer Science Faculty Publications

Federated Learning (FL) often suffers from unstable convergence and reduced robustness under non-IID data and malicious attacks. In this paper, we present FedPIDAvg_tuned, a control-theoretic aggregation framework for improving stability and adversarial robustness in FL. In particular, our approach combines a server-side Proportional–Integral–Derivative (PID) controller with Bayesian optimization to tune controller gains for different data and attack conditions. The PID controller regulates global model updates through feedback on loss dynamics, providing adaptive scaling that improves stability and convergence. The tuned gains are applied within a trust-weighted trimmed-mean mechanism to remove adversarial or outlier updates. Using the Flower framework, we evaluate …

Redefining Educational Technology: A Critical Collaborative Inquiry, Aras Bozkurt, Helen Crompton, Robert Farrow, Agnes Kukulska-Hulme, Jon Dron, Richard West, Agnieszka (Aga) Palalas, Matth Bower, Junhong Xiao, Ahmed Tlili, Danah Henriksen, Angelica Pazurek, Henk Huijser, Thomas K. F. Chiu, Petar Jandrić, Katy Jordan, John Curry, Royce Kimmons, Mutlu Cukurova, Thomas Reeves, Gwo-Jen Hwang, Peter Shea, Jason Lodge, Martin Weller, Davy Ng, Tutaleni Iita Asino

STEMPS Faculty Publications

Educational technologists have not settled on a fixed definition of the field and likely never will. However, attempting to define the field helps to understand the epistemological meanings that shape what the field sees, values, and considers worth pursuing. Through a critical historical review spanning over a century, alongside theoretical engagement with the concepts of entanglement and distributed agency, this paper identifies three key insufficiencies in current educational technology frameworks. These are the persistence of an instrumental-facilitative paradigm that treats technology as a resource deployed by human agents; the theoretical dissolution of the pedagogy-technology dichotomy that existing definitions have not …

Ai-Powered Lawyering: Ai Reasoning Models, Retrieval Augmented Generation, And The Future Of Legal Practice, Daniel Schwarcz, Sam Manning, J. J. Prescott, Patrick Barry, David R. Cleveland, Beverly Rich

Articles

Generative AI is set to transform the legal profession, though its most promising uses and ultimate effects are still unclear. While AI models like GPT-4 improve efficiency, they can also “hallucinate” and may undermine legal judgment, particularly in complex tasks typically handled by skilled lawyers. This article examines two emerging AI innovations that may mitigate these concerns: Retrieval Augmented Generation (RAG), which grounds AI-powered analysis in legal sources, and AI reasoning models, which structure complex reasoning before generating output. We conduct the first randomized controlled trial assessing these technologies, assigning upper-level law students to complete legal tasks using a RAG-powered …

Javatutor: An Integrated Llm Framework For Structured Programming Tutoring And Problem-Solving In Java, Amey Makarand Dhongade

Master's Projects

Large Language Models (LLMs) are being adopted more actively to assist in pro- gramming tasks, but much of the available AI code assistants devices are encouraging the rapid generation of solutions, as opposed to assisting students in developing an understanding. As a result, they can give the students correct code without teaching on how to judge a problem and how to build solutions or how to debug their programs themselves. This study describes JavaTutor, an LLM-based tutoring system that is used to assist learners in the learning process in a three-step process of understand- ing, generating a solution, and refining …

Safe-R2r: A Safety-Aware And Budget-Optimized Retrieval Controller For Rag Systems, Mandar Sunil Gondane

Master's Projects

Retrieval-Augmented Generation (RAG) improves factual grounding in large language models by incorporating external evidence during inference. However, most RAG systems rely on fixed retrieval strategies that ignore query difficulty, computational cost, and prediction uncertainty. This project introduces SAFE-R2R (Safety-Aware and Budget-Optimized Reason-to-Retrieve), a retrieval controller that treats retrieval as a query-dependent decision problem. SAFE-R2R organizes retrieval into a multi-rung ladder ranging from no retrieval to deeper retrieval with reranking. At each rung, the system generates an answer, computes reliability signals, and combines them into a risk score used within a conformal calibration framework to decide whether to accept the answer …

Multimodal Emotion Detection System, Shubhankar Sameer Munshi

Master's Projects

Trying to understand emotion from speech is a problem that is present in human computer interaction. Nevertheless, there are still some shortcomings in current SER methods. Text-based systems may miss vital vocal cues, such as sarcasm, tone changes, and delivery. On the other hand, purely audio-based systems are prone to noise and unstable acoustic features. The combination of linguistic and acoustic features in multimodal approaches partially solves this problem, but many existing approaches use inflexible multimodal fusion techniques that cannot adjust their behaviors according to the quality of input signals. In this work, we propose a multimodal approach based on …

Modeling Sealed Deck Construction In Collectible Card Games Using Learning-To-Rank Approach, Michael Dinh Nguyen

Master's Projects

Artificial intelligence has demonstrated strong performance in complex decision-making domains such as chess and Go, motivating research into its application for games with even richer rules and combinatorial complexity. In collectible card games like Magic: The Gathering, deck construction from a constrained card pool is a critical and challenging task that requires evaluating card strength, synergy, and resource balance. This project explores whether machine learning, specifically learning-to-rank (LTR), can effectively model these human decision processes to construct competitive decks in a sealed format. The results found here can also be applicable to other areas of note, such as sports drafting …

Stepwise Sudoku Reasoning Training Using Transformers, Lok Man Chu

Master's Projects

Sudoku is a constraint satisfaction problem that serves as a testbed for studying reasoning and stepwise deduction. While many solvers can produce correct solutions, they often fail to generate human interpretable sequences of logically consistent steps. This study investigates whether a transformer trained on stepwise deduction traces can learn to solve Sudoku puzzles through sequential, logically deducible moves. Results show that solve accuracy improves significantly with more training data, reaching approximately 78% at 1500k samples. Stepwise analysis indicates that the model effectively learns simple strategies, achieving near-perfect performance. However, performance on more complex strategies remains limited. Overall, while the model …

Multi-Agent Path Planning And Optimization Using Q-Learning, Chirag Rudresh

Master's Projects

Robot navigation in a multi-agent setting requires a balance between safety and efficiency, especially in dense environments. In these two-dimensional spaces, the scope for geometric errors is much less and could lead to collisions or immobility. This project proposes to address the navigation task using a two-phase path planning pipeline that combines reinforcement learning and convex optimization in a scalable and robust manner. The first phase consists of generating diverse collision-free paths using a Q-learning agent that is trained on a visibility graph representation of the environment. The discretization of the environment using waypoint-based graphs allows the agent to train …

Real-Time Isolated Asl Recognition: Evaluating Spatial-Temporal Networks And Multimodal Llms, Raga Mouni Batchu

West Chester University Graduate Theses, Dissertations, and Final Projects

This thesis investigates the deployment of high-accuracy Isolated ASL Recognition (ISLR) in resource-constrained edge environments. We train a lightweight Spatio-Temporal Attention Network (SSTAN,∼2.7 M parameters,∼10 MB) on the WLASL-100 benchmark, achieving 75.25% Top-1 and 88.24% Top-5 accuracy with 139 ms CPU-only inference. A systematic comparison against frontier multimodal LLMs (Gemini 3 Flash, Gemini 3.1 Pro, Qwen 3 VL) shows SSTAN outperforms the best LLM baseline by∼1.85×in accuracy while being 22–230×faster and up to 40×cheaper annually. The LLMs’ core limitation is a lack of fine-grained temporal perception; they impose English-language semantic priors rather than learning the articulatory distinctions that define ASL …

Polysemicolon; Novice Programmers And Java Keywords, Briana C. Bettin

Michigan Tech Publications

Most industrial programming languages leverage the English language for reserved keywords – words which a program compiler recognizes as specific execution commands. The divide between expert and novice programmers showcases an intriguing middle-ground by which the polysemy of many keywords becomes revealed. This essay explores a sampling of the multitude of keyword interpretations that a novice programmer may derive from the Java language’s syntactic style and keywords specifically, and how the polysemy of both “English” and “code” meanings to these terms affects the novice-expert programmer transition.The transition from novice to expert, and the mapping of the career of metaphor to …

The Devil Is In The Det[Ai]Ls: Ai Agents, Ghost Students, And The Crisis Of Verified Presence In An Agentic Ai World, Aras Bozkurt, Helen Crompton, Caroline Fell Kurban

STEMPS Faculty Publications

The transition from reactive Generative Artificial Intelligence (GenAI) to agentic AI systems marks a categorical shift in digital education, moving beyond simple content generation to goal-oriented, autonomous execution. This paper explores the emergence of the “ghost student”: a digital surrogate created by the coupling of Large Language Models (the “mind”) and agentic AI browsers (the “body”). These entities are capable of navigating Learning Management Systems (LMS), engaging with content, and completing assessments with human-like mimicry, often rendering the actual learner’s presence optional. We argue that this phenomenon creates a verification gap that traditional proctoring and detection tools are structurally unable …

Instructional Designers' Reflections On Generative Ai Use For Scenario-Based And Performance-Centered Learning: A Collective Autoethnography, Nour El Houda Maache, Dan V. Dao, Jiyoon Jung, Jayanth Nadheri, Nari Kim, Chikezie Ozuzu, Xinyue Ren

STEMPS Faculty Publications

This study examines how instructional designer-instructors (IDIs) use and evaluate generative artificial intelligence (GenAI) when designing scenario-based and performance-centered authentic learning in higher education. Using a collective autoethnography (CAE) approach, the study draws on semi-structured interviews and reflective narratives from five IDIs with varied professional experience. Findings indicate that GenAI enhanced design capacity by accelerating scenario development, translating complex content, and supporting scenario-based and performance-based task construction. At the same time, participants reported limitations related to contextual misalignment, output unreliability, and the cognitive demands of prompt refinement. Across cases, effective integration depended on sustained human oversight, disciplinary judgment, and ethical …

Chatgpt In Secondary English Language Learning: Educators' And Students' Literacy, Perceptions, And Experiences, Victoria Brenes, Sierra Outerbridge, Xinyue Ren

STEMPS Faculty Publications

The increasing use of generative artificial intelligence (GenAI) has shown the potential of transforming teaching and learning practices in various educational settings, such as in English language learning (ELL). As English language learners (ELLs) often experience many challenges and barriers in schools in the United States, it is urgent to leverage the educational affordances of GenAI in fostering the effectiveness of ELL. Given the limited research investigating GenAI adoption, especially ChatGPT literacy within K-12 ELL, this convergent mixed methods research aims to investigate students' and teachers’ perceptions of using ChatGPT and their ChatGPT literacy in secondary ELL contexts. We will …

A Comparison Of Conversational Chatbots And The Internet For Consumer Information Search, Wondwesen Tafesse, Yoseph Mamo

STEMPS Faculty Publications

This study compares consumer perceptions of conversational chatbots and the internet for information search. While the internet is a mature platform, conversational chatbots represent an emerging technology, and insight into how consumers view them in relation to the internet for information search is lacking. Drawing on the information source utility perspective, the study builds a comparative model based on four key dimensions: information currency, information customisation, information trustworthiness, and media richness. Additionally, the study investigates consumers’ prior experience with conversational chatbots as a moderating factor. Data was collected from 191 respondents recruited through MTurk. Paired sample t-tests assessed mean differences …

Automatic Generation Of Introductory Programming Exercises With Large Language Models, Nguyen Binh Duong Ta, Hua Gia Phuc Nguyen, Gottipati Swapna

Research Collection School Of Computing and Information Systems

Despite recent advances in code generation made possible by large language models (LLMs), programming is still an essential skill that computing students need to master now and in the foreseeable future. In learning programming, frequent practices with exercises set at an appropriate difficulty and knowledge level is of crucial importance for students. However, it’s not a trivial task for instructors to create many good quality exercises customized for each student. Programming problems found on Internet sources such as LeetCode are mostly too challenging for novice programmers with no prior coding knowledge. Recent work in AI-enabled education has been leveraging LLMs …

Can We Read Ai’S Mind? A Quest For Transparency, Santhosh Kumar Ravindran, Estera Kot, Fiona Fui-Hoon Nah

Research Collection School Of Computing and Information Systems

Although Artificial Intelligence (AI) systems are playing an increasing role in critical domains such as healthcare, finance, and autonomous systems, their decision-making processes remain largely opaque. This paper examines the challenges of AI transparency, addressing the “black box” problem using Explainable AI (XAI) techniques such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). It also examines the ethical, regulatory, and societal implications of AI opacity and proposes a Comprehensive AI Observability (CAO) Framework that integrates deep explainability, provenance tracking, and real-time monitoring to enhance AI accountability. By bridging technical solutions with governance structures, this research emphasizes the …

Design Principles For Customer-Engaging Digital Service Systems: An Action Research Study, Keng Leng Siau, Xiaofeng Chen, Xin Tan

Research Collection School Of Computing and Information Systems

Digital services represent a business approach employed by organizations to operate in the digital environment. However, systematic development guidelines for developing quality digital service systems are lacking in the literature. The authors identified four general challenges for developing and implementing customer-engaging digital service systems (CEDSS). By employing the method of canonical action research in a digital service system project, they derived 10 design principles for developing high-quality CEDSS. They empirically evaluated the design principles in the development project and through follow-up focus group sessions. The design principles provide applicable and actionable guidelines for the development of CEDSS.

Towards Secure Program Partitioning For Smart Contracts With Llm’S In-Context Learning, Ye Liu, Yuqing Niu, Chengyan Ma, Ruidong Han, Wei Ma, Yi Li, Debin Gao, David Lo

Research Collection School Of Computing and Information Systems

Smart contracts are highly susceptible to manipulation attacks due to the leakage of sensitive information. Addressing manipulation vulnerabilities is particularly challenging because they stem from inherent data confidentiality issues rather than straightforward implementation bugs. To tackle this by preventing sensitive information leakage, we present PARTITIONGPT, the first LLM-driven approach that combines static analysis with the in-context learning capabilities of large language models (LLMs) to partition smart contracts into critical (privileged) and normal codebases, guided by a few annotated sensitive data variables. We evaluated PARTITIONGPT on 18 annotated smart contracts containing 99 sensitive functions. The results demonstrate that PARTITIONGPT successfully generates …

Llamoco: Instruction Tuning Of Large Language Models For Optimization Code Generation, Zeyuan Ma, Yue-Jiao Gong, Hongshu Guo, Jiacheng Chen, Yining Ma, Zhiguang Cao

Research Collection School Of Computing and Information Systems

Recently, combining the strength of large language models (LLMs) and Evolutionary Computation (EC) has shown promising results for addressing optimization problems. It typically involves either iterative next-step solution seeking or directly prompting LLMs to generate critical optimization codes. However, these methods often suffer from low computational efficiency, high sensitivity to prompt design, and a lack of domain-specific knowledge. We introduce LLaMoCo, the first instruction-tuning framework designed to adapt LLMs for solving optimization problems in a code-to-code manner. LLaMoCo features a comprehensive instruction set that includes code-style problem descriptions as input prompts and robust optimization codes from expert EC optimizers as …

Energy-Efficient Security For Narrowband Iot Using Blockchain And Ep-Cumac, Hafizullah Kakar

UNF Graduate Theses and Dissertations

The Narrowband Internet of Things (NB-IoT) continues to expand but faces challenges such as cryptographic overhead and energy consumption. Security frameworks such as blockchain and Energy-Performance Cumulative Message Authentication Codes (EP-CuMAC) rely heavily on SHA-256, which is not optimized for energy-limited devices.

This work unifies two complementary approaches, a hybrid blockchain-based NB-IoT framework and an EP-CuMAC-based framework, by engineering their cryptographic core with an Energy Complexity Model-optimized SHA-256 (ECM-SHA256). ECM applies parallel memory-bank mapping and block-level access optimization to reduce redundant power usage while preserving algorithmic integrity.

Experimental evaluation on identical Intel DDR3 systems using pyRAPL shows energy savings of …

From Authorization To Loss: A Blockchain Forensic Analysis Of Transaction-Level Mechanisms In Cryptocurrency Airdrop Scams, Chanwoo Shin, Kyung-Shick Choi

International Journal of Cybersecurity Intelligence & Cybercrime

Cryptocurrency airdrop scams have emerged as a rapidly growing form of cyber-enabled financial crime, yet remain underexplored in empirical research. This study examines how transaction-level mechanisms and offender strategies influence variation in monetary loss in airdrop scam incidents. Grounded in Cyber-Routine Activities Theory (Cyber-RAT), the study conceptualizes financial harm as occurring within decentralized environments where users’ online behaviors, particularly transaction authorization, intersect with limited digital capable guardianship. Data were drawn from 112 validated airdrop scam cases reported on Chainabuse.com between January and December 2025. Blockchain forensic analysis using Breadcrumbs was conducted to reconstruct transaction pathways, identify exchange interactions, and detect …

Unsafe2safe: Controllable Image Anonymization For Downstream Utility, Minh Dinh

Dartmouth College Master’s Theses

Large-scale image datasets frequently contain identifiable or sensitive content, raising privacy risks when training models that may memorize and leak such information. We present Unsafe2Safe, a fully automated pipeline that detects privacy-prone images and rewrites only their sensitive regions using multimodally guided diffusion editing. Unsafe2Safe operates in two stages. Stage 1 uses a vision--language model to (i) inspect images for privacy risks, (ii) generate paired private and public captions that respectively include and omit sensitive attributes, and (iii) prompt a large language model to produce structured, identity-neutral edit instructions conditioned on the public caption. Stage 2 employs instruction-driven diffusion editors …

Lens: Llm-Enabled Narrative Synthesis For Mental Health By Aligning Multimodal Sensing With Language Models, Wenxuan Xu

Dartmouth College Master’s Theses

Multimodal health sensing offers rich behavioral signals for assessing mental health, yet translating these numerical time-series measurements into natural language remains challenging. Current LLMs cannot natively ingest long-duration sensor streams, and paired sensor–text datasets are scarce. To address these challenges, we introduce LENS, a framework that aligns multimodal sensing data with language models to generate clinically grounded mental-health narratives. LENS first constructs a large-scale dataset by transforming Ecological Momentary Assessment (EMA) responses related to depression and anxiety symptoms into natural-language descriptions, yielding over 100,000 sensor–text QA pairs from 258 participants. To enable native time-series integration, we train a patch-level encoder …

From Physical Correlation To Emotional Connection: The Role Of Passive Haptics On Empathy In Virtual Reality, Jemely Robles

Dartmouth College Master’s Theses

Virtual reality is increasingly explored as a tool for cultivating empathy, and haptic feedback as a tool for enhancing immersion. This paper investigates the effects of combining the two.  Fifty-two participants experienced a custom-built VR scene in which a character was shown packing up a room. Participants were assigned to either a haptic condition, receiving passive haptic feedback corresponding to the character's actions, or a non-haptic control condition that did not receive any haptic input. Trait empathy was measured beforehand, and state empathy and engagement were measured after the experience. Thematic analysis was conducted on post-study interviews, and headset recordings …