Artificial Intelligence and Robotics Commons^™

Automatically Constructed Preference Pairs For Chain-Of-Thought: Consistency Gains With Accuracy Tradeoffs, Cameron Scolari, Lanyu Shang

Honors Thesis

We investigate preference optimization over chain-of-thought (CoT) reasoning using automatically constructed preference signals derived from the accuracy and internal consistency of a model. Our results show that framing reasoning as a preference learning problem improves both the accuracy of the final answer and the structure of the model outputs. We observe a non-monotonic relationship between performance and the Direct Preference Optimization (DPO) scaling parameter β, where moderate values maximize accuracy while lower values improve stability, highlighting a tradeoff between optimization strength and reliable generation. We further identify a tradeoff between reasoning consistency and accuracy. Increasing the consistency weight improves agreement …

Ai Shell, Jacqueline Roebuck Sakho

Books

An AI Configuration Shell to train and deploy an Argumentation Coach. Presented and distributed to the UNC Commons Collective in May 2026.

Synthergy: Social Deduction And Deception In Llm-Powered Agents, Lauren Campbell, Andrew Forney

Honors Thesis

Synthergy is an online social deduction game designed to enable comparative analysis of how large language model-powered agents engage in social deduction and deception under conditions of asymmetric information. Inspired by social deduction games such as Town of Salem, Throne of Lies, and Mafia, the game consists of two factions, Harmony and Discord, to which agents are secretly assigned. Agents must infer others’ affiliations through dialogue, in-game abilities, and voting behavior. To evaluate agent behavior, we conducted 100 simulated games across six agent types: a random baseline agent (RandomSynth), an LLM-based agent (Synth), a chain-of-thought agent (CoT Synth), a Bayesian …

Llms In Compiler Construction, Raffi Khatchadourian

Open Educational Resources

These lecture slides survey the use of large language models (LLMs) in compiler construction for a graduate compiler course (CSc 81010). They situate LLMs across the compiler pipeline and examine representative work: foundation models trained on LLVM IR and assembly (Meta's LLM Compiler), LLM-driven code optimization, binary decompilation (LLM4Decompile), and LLM-assisted automated refactoring—alongside the challenges of applying probabilistic models to tasks that demand correctness. The slides are a self-contained HTML (W3C Slidy) deck with editable Pandoc Markdown source. Part of a two-session unit on advanced compiler topics; see also "Deep Learning Compilers."

Deep Learning Compilers, Raffi Khatchadourian

Open Educational Resources

These lecture slides introduce deep learning compilers for a graduate compiler-construction course (CSc 81010). Building on the classical compiler pipeline, they show how modern machine-learning systems compile tensor programs: static tensor and type analysis (illustrated by a WALA/Ariadne-based refactoring of imperative TensorFlow code to graph mode), MLIR-based end-to-end compilation with IREE, and the PyTorch 2.x stack—TorchDynamo graph capture, AOTAutograd, PrimTorch operator decomposition, and TorchInductor lowering to Triton (GPU) and C++/OpenMP (CPU). The slides are a self-contained HTML (W3C Slidy) deck with editable Pandoc Markdown source. Part of a two-session unit on advanced compiler topics; see also "LLMs in Compiler Construction."

Shaped By The Feed: Big Data, Identity And Power, Ana Paula Andrada

COD Library Student Research and Award Symposium

I explored how big data and algorithms shape the way we think and interact at an individual and societal level. This started from things I kept noticing in everyday life. Through my research, I learned how the personalization and predictability in feeds can reinforce beliefs, limit our critical thinking, deepen polarization and threaten our freedom.

Faculty Sponsor:  Professor Aleisha Balestri

Dynamic Trust Calibration, Bruno Miranda Henrique

Dartmouth College Ph.D Dissertations

Trust calibration between humans and Artificial Intelligence (AI) is crucial for optimal decision-making in collaborative settings. Excessive trust can lead users to accept AI-generated outputs without question, overlooking critical flaws, while insufficient trust may result in disregarding valuable insights from AI systems, hindering performance. Despite its importance, there is currently no definitive and objective method for measuring trust calibration between humans and AI. Current approaches lack standardization and consistent metrics that can be broadly applied across various contexts, and they don’t distinguish between the formation of opinions and subsequent human decisions. This thesis brings a novel and objective method for …

Meeting The Moment With Ai-Employer Informed Education, Brent Terwilliger, John Faraca

Publications

As artificial intelligence transforms aviation, aerospace, and autonomy-related sectors, higher education must adapt to meet evolving workforce demands. This session shares emerging findings from a nationwide study led by Embry-Riddle Aeronautical University, focused on employer perceptions of AI adoption, responsible use, and workforce preparedness in domains including uncrewed systems, space systems, robotics, and advanced air mobility. Based on a structured survey and follow-up interviews, the presentation explores how organizations are using AI tools, from generative platforms to enterprise systems, and defining effective and inappropriate use in operational contexts. Participants will gain insight into critical concerns (e.g., data privacy, compliance, security, …

Labeling And Describing Objects In A Photogrammetry-Based 3d Environment Using Computer Vision And Artificial Intelligence, Colin Donald Moschella

Capstone Projects

Cataloging and digitizing the objects inside a building manually is a task that is often impractical at scale. This project therefore automates the process, using a custom-made system. Using a photogrammetry-based 3D reconstruction of a room, this system is applied to sequences of 2D images used to make the 3D models. The system applies object detection, image segmentation, and image-text models to identify and describe objects, using CNN based models such as YOLO and OpenCLIP. Each analyzed object is then stored in a structured database with spatial coordinates from the 3D scanning, descriptive attributes from the image-text models, and other …

The One-Shot Ceiling: Comparing Rag And Fine-Tuning Architectures For Ai-Assisted Math Mentoring, Michael J. Cummins

Computer Science Honors Papers

Providing individualized feedback to math students is a resource-intensive bottleneck in STEM education. We present Mentir-AI, a tool designed to elevate teacher capacity by generating high-quality mathematical feedback using the Mathforum's "Problem of the Week" archive. By analyzing a corpus of nearly one million interactions, we compare the efficacy of Retrieval-Augmented Generation (RAG) and Fine-Tuning (FT) architectures. This study details the development of an automated grading pipeline, the evolution of a multi-component system prompt, and the implementation of an automated mentor grading system in AI-led evaluation. While Fine-Tuning demonstrates superior instructional judgement, our results identify persistent failure modes in mathematical …

How Ai Governance Differs Between Regime Type, Jackson T. Guillou

Honors Theses

This thesis will examine how artificial intelligence (AI) regulation and development differ across political regime types, arguing that governance outcomes are fundamentally shaped by institutional political structures. This thesis will draw on comparative frameworks analyzing democratic and authoritarian systems. This thesis will also include case studies of the European Union, the United States of America, China, and Russia. The study finds that democratic regimes tend to emphasize transparency, accountability, and rights-based regulation of AI. This often results in slow regulation of AI, but it is more ethically legitimate and constrained. In contrast, authoritarian regimes prioritize centralized control, strategic coordination, and …

Stridevision: Automated Detection Of Running Form Deviations From 2d Pose Estimation And Machine Learning, Paulina Eguibar Ortega

Honors Theses

Running gait analysis plays a critical role in injury prevention and performance optimization, however, existing approaches often rely on specialized laboratory equipment or wearable sensors with limited interpretability. Recent advances in computer vision, particularly 2D human pose estimation, enable markerless motion analysis from standard video. However, progress remains constrained by the lack of publicly available datasets designed for running form analysis.

In this work, we introduce a preliminary dataset and benchmark for stride-level running gait analysis. The dataset consists of 73 treadmill running videos from 15 participants with varying experience levels, annotated with over 4,600 stride-level labels across multiple biomechanical …

Private Assistant Agent, Bao Le '26

Senior Scholarly and Creative Symposium

Our brain has limited fuel: every time we make a decision, it costs fuel. A Personal Artificial Assistant will help preserve that fuel by handling secretarial tasks and scheduling a daily timetable for its user. Personal Assistant Agent (PAA) is an AI assistant running as a text-based Windows desktop application to help users in scheduling and give reminders like a secretary. Users interact with the PAA conversationally. This project is a system of reasoning and a modular memory pipeline that makes any Large Language Model (LLM) behave like a personal assistant. The Language model is isolated inside a function to …

Psychiatry Meets Ai: Are Residents Ready?, Jacob De Castro, John Case

Rowan-Virtua Research Day

Artificial intelligence (AI) use is increasing in healthcare, but psychiatry residency training remains unstructured. In a 20-resident pilot survey, AI was frequently used for literature review and clinical support, with limited confidence and institutional guidance. We found most residents desired formal training and would use AI more if institutionally supported. Findings highlight a gap between rapid adoption and structured education.

Ai-Augmented Digital Auscultation For Point-Of-Care Screening Of Valvular Heart Disease: A Systematic Review And Meta-Analysis, Harshal Parmar, Akhila Archakam, Nikhila Archakam, Wesley Kim, Eduard Koman Md

Rowan-Virtua Research Day

Background: Valvular heart disease (VHD) affects >10% of adults aged 75+ yet remains underdiagnosed when asymptomatic due to declining auscultatory proficiency. AI-augmented digital auscultation offers a point-of-care screening solution, though no meta-analysis has pooled diagnostic accuracy across VHD subtypes in adults using echocardiography as the reference.

Methods: A systematic review and meta-analysis were conducted per PRISMA guidelines. PubMed, Embase, Cochrane, IEEE Xplore, and Scopus were searched without date restriction. Studies applying AI or machine learning to digital auscultation or phonocardiography for VHD classification in adults with echocardiographic reference and patient-level diagnostic metrics were included. Studies using only public datasets, pediatric …

Developing A Comprehensive Resource Guide For Caregivers Within The Area Agency On Aging, Aiden R. Murphy

Public Health Capstone Projects

This project developed a new universal caregiver resource guide for caregivers within the Area Agency on Aging, in order to improve resource navigation and workflow efficiency. Resources were collected, verified, and organized into a new, streamlined guide via the ARIA chatbot, covering multiple needs. Caregiver resources were collected and verified by the capstone student and mentor, Michael Kroeker, and organized into a centralized knowledge base within the ARIA chatbot. A mixed methods evaluation was conducted utilizing a 5-point Likert scale with three quantitative questions and one open-ended qualitative question. The data was given to the SeniorLine staff, who wanted to …

What History Shows: A Structural Account Of Fixed-Point Theory Formation, Griselda Poe

Publications and Research

Theory generation has long been subsumed under categories such as creativity, genius, and innovation. These categories do not distinguish assembly-based conceptual synthesis from fixed-point theory generation.

This paper makes that distinction explicit. The termination condition of assembly is external: data, citation, endorsement, usability. The termination condition of fixed-point theory generation is internal: consistency with internally held constraints, resolution of structural contradiction. The two operate under different processing conditions.

The historical record confirms this distinction. What Darwin, Einstein, Spinoza, and Kant produced was not assembly. Their processes involved unresolved branch retention and decomposition necessity, arriving at internally constrained fixed-point termination. Freud …

Predicting And Decoding Allosteric Binding Sites Using Protein Language Models And Structure-Based Machine Learning: An Energy Landscape-Guided Explainable Ai Framework, Kamila Riedlová, Vít Skrhák, William G. Gatlin, Max Ludwick, Lucas Turano, Marian Novotný, David Hoksza, Gennady M. Verkhivker

Mathematics, Physics, and Computer Science Faculty Articles and Research

Computational prediction of allosteric binding sites in protein structures remains a persistent challenge, as these regulatory pockets evade detection by both sequence-based and structure-based algorithms. Both computational and physical origins of this predictive asymmetry remain insufficiently understood. In this study, we systematically examine the determinants of binding site predictability using a dual framework that integrates a fine-tuned protein language model and the structure-based method P2Rank as complementary tools probing a diverse data set of 453 human kinases, together with a physics-based interpretability layer derived from energy landscape frustration analysis. Both predictors exhibit a sharp and reproducible dichotomy on protein kinases, …

Propasafe-Hybrid: A Text-Based Hybrid Propaganda Detection Tool, Thomas Kimmeth, Avijit Roy, Vivek Sharma

Publications and Research

Propagandistic content increasingly circulates through online news and social media, where readers often encounter it with limited scrutiny, highlighting the need for reliable and fine-grained detection. This paper introduces Propasafe-Hybrid, a sentence-level system that integrates a fine-tuned transformer classifier with LLM-based technique classification to identify, label, and explain specific propaganda strategies. The pipeline generates actionable outputs, including highlighted sentences, technique assignments, and concise rationales, so users can immediately understand why a sentence was flagged and how each label was determined. To control inference cost, Propasafe-Hybrid employs a cost-aware pre-filtering stage that forwards only high-likelihood sentences to LLMs, reducing token usage …

Toward Vehicle-Agnostic Driving Signatures For Cognitive Impairment Prediction From Naturalistic Driving Data, Aadarsha Gopala Reddy

McKelvey School of Engineering Graduate Student Theses & Dissertations

This thesis studies whether naturalistic driving data can help predict binary Clinical Dementia Rating (CDR) status while accounting for differences across vehicles. The final analytic dataset comprised 26,968 participant-weeks from 304 participants. Weekly driving features were derived from real-world telematics data and combined with four demographic covariates. Primary model comparisons used leave-one-participant-out (LOGO) cross-validation, with one individual held out at a time and pooled participant-level metrics used as the main reporting surface.

The main comparison includes six model families evaluated on the same dataset under a shared LOGO framework. Performance remained modest overall. GRU-DANN had the highest participant-level ROC AUC …

Improving Semantic Precision In Text-To-Image Diffusion Models Via Latent-Space Optimization And Semantically-Parsed Evaluation, Mohammad Rouie Miab

McKelvey School of Engineering Graduate Student Theses & Dissertations

Text-to-image diffusion models can produce visually impressive images from natural-language prompts, but they often fail to satisfy the detailed semantic constraints expressed in compositional prompts. Typical failure modes include omitted objects, merged entities, incorrect quantities, incorrect attribute binding, and leakage of one entity's attributes onto another. This thesis studies the problem of semantic precision in text-to-image generation: how faithfully a generated image satisfies the structured meaning of its prompt.   The thesis makes two linked contributions. First, it presents a training-free inference-time refinement method for diffusion-based image generation. The method operates directly in latent space during denoising and uses noun-phrase-aware cross-attention …

Interpreting American Sign Language: A Literature Review Of Assistive Technologies, Natalie Louise Paradiso, Emma Grace Kochenderfer

Student Scholar Symposium Abstracts and Posters

American Sign Language (ASL) is a visually elaborate, spatially oriented linguistic methodology that relies on combinations of hand movements, body positioning, facial expressions, and motion/spatial perception, aspects of which make interpretation difficult for automated machine recognition. Current assistive technology approaches to ASL interpretation are generally within the categories of computer vision models (including deep learning, multi-focus image fusion, and keypoint tracking) and wearable, multimodal/sensor-based approaches (such as smart glasses and inertial-sensor gloves). Within controlled environments, computer vision models perform well. However, when applied to conditions such as non-manual signs/features, signer variability, and rapid assimilation, they falter in processing all aspects …

Developing Narrative-Based Stem Learning Tool For K-6 Visually Impaired Students, Daniel Tsivkovski, Dylan Ravel, Jeffrey Kraskouskas, Brandon Foley, Maryam Etezad, Franceli Cibrian, Rajeev Joshi, Ariel Han

Student Scholar Symposium Abstracts and Posters

This research develops a free, accessible web application that enables K-6 students who are blind or visually impaired (BVI) to learn STEM concepts using refreshable braille displays. Currently, most online learning tools are not designed for BVI students, creating a significant educational barrier.

The application interfaces with commercial braille displays and uses narrative-based learning to make STEM content approachable and engaging. By presenting material as personalized interactive stories generated with the help of Artificial Intellligence (AI), students can connect with concepts while developing braille reading skills. The curriculum design prioritizes accessibility through the Accessible Rich Internet Applications (ARIA) standards and …

Security Assessment Of A Machine Learning Approach To Generate And Validate Digital Signatures, Juan Ortiz Couder

Doctoral Dissertations and Master's Theses

Cybersecurity has become a global concern as cyber-attacks have become more common, and the cost of the damage caused by them continues to increase. There are several approaches to improve the cyber security of systems such as Digital Signatures, hashing, watermarking, and encryption among others. Digital Signatures are a cryptographic technique used to verify the authenticity and integrity of digital messages or documents. Digital Signatures use a combination of hashing and public-private key encryption to verify the authenticity and integrity of videos, just as they are used for documents and messages. As a result of using a combination of other …

Machine Learning For Handwritten Character Recognition, Hannah Freitag

Honors Capstones

Handwritten character recognition remains a challenging problem in machine learning due to the high variability of handwriting across individuals and the visual similarity between certain character classes. This project explores whether Singular Value Decomposition (SVD)-based dimensionality reduction can serve as an effective preprocessing step for a fully connected neural network trained on the EMNIST Balanced dataset, a 47-class benchmark of handwritten digits and letters. By projecting 784- dimensional pixel inputs onto the top 70 principal components, approximately 90% of the total variance is preserved while reducing input dimensionality by 91%. The resulting SVD-based model achieves approximately 94% test accuracy, outperforming …

Propasafe-Hybrid: A Text-Based Hybrid Propaganda Detection Tool (Ila 2026 Presentation), Thomas Kimmeth, Avijit Roy, Vivek Sharma

Publications and Research

This presentation introduces Propasafe-Hybrid, a hybrid system for sentence-level propaganda detection that combines offline transformer-based classification with selective large language model (LLM) explainability. The system employs a two-stage pipeline in which a local BERT-based classifier evaluates all input text and filters non-propagandistic content, while only high-confidence candidates are forwarded to an LLM for rhetorical technique labeling and explanation. This design enables cost-aware, privacy-conscious, and scalable analysis by reducing unnecessary reliance on external models.

Propasafe-Hybrid identifies propagandistic techniques such as loaded language, obfuscation, and appeal to fear, and generates concise natural language rationales that make these techniques interpretable to users. By …

A Survey On Knowledge-Enhanced Healthcare Question Answering Systems, Junnan Su, Pu Han, Jianxiang Wei

Journal of Scientific Information Research

[Purpose/significance] This paper aims to review the research progress and applications of knowledge enhancement techniques in healthcare question answering systems, in response to the limitations of traditional systems in knowledge representation and reasoning, as well as challenges faced by current large language model-based systems, such as insufficient domain knowledge, privacy concerns, and hallucination. The review provides a systematic reference for improving the precision and knowledge reliability of such systems. [Process/method] Focusing on knowledge enhancement strategies, this paper firstly outlines their fundamental concepts and overall framework. The strategies are then categorized into explicit and implicit types, with an analysis of their …

Research On Text Translation Model Based On Large Language Model And Knowledge Enhancement Framework, Chuanming Yu, Haoxuan Li

Journal of Scientific Information Research

[Purpose/significance] This paper aims to improve the translation quality of large language models and effectively alleviate the translation illusion problem, thereby enhancing cross-linguistic information retrieval capabilities. [Method/process] A translation generation method based on a knowledge enhancement framework is proposed. This framework optimizes the translation process from multiple dimensions, such as style, focus, and cultural adaptability, by combining external knowledge provided by the translation context building module and the knowledge base building and retrieval module, and then utilizing the guidance of the text attention module. [Result/conclusion] Experimental results show that the proposed method effectively enhances model performance. Specifically, on the WikiLingua, …

Sequential Robustness In Adversarial Reinforcement Learning, Roman Lok-Ming Belaire

Dissertations and Theses Collection (Open Access)

My goal is to build autonomous systems that expand the reach of human capability in challenging domains such as undersea and space exploration, disaster response, and large-scale infrastructure. In everyday settings, these systems will increasingly appear in safety-critical applications such as autonomous driving, robotics, and industrial manufacturing. A central requirement for these systems is the ability to operate reliably under uncertainty, particularly when the environment behaves in unanticipated ways.

The robust handling of unforeseen environment dynamics is therefore a technical cornerstone of autonomous decision-making; Adversarial attacks provide a useful and principled lens through which to study this problem. Adversarial \textit{robustness}, …

Perspectives On Interpretability For Neural Text Representations, Jia Peng Lim

Dissertations and Theses Collection (Open Access)

In this dissertation, we investigate interpretability in the three elements of learning neural text representations: inputs, passed into models, to produce probabilistic outputs. We emphasise perspectives as we present alternative novel methods to mine and organise meaning in this work.

Models. We initiate our investigation by examining Neural Topic Models (NTM), proposing an alternate angle of interpreting its word-topic distribution, producing better topic representations for interpretation. Our method maps the problem of finding these better interpretations to classical NP-hard graph problems, enabling examination of topic distributions in a composite manner. Next, we apply our previous findings to extract interpretations from …