Computer Science
Personalized, Context-Driven Agentic Systems for Academic Advising
Presenter: Sean B. Collins
Group Members: Luca Antonio Silver, Noel Mensah
Faculty Sponsor: Hao Loi
School: Quinsigamond Community College
Research Area: Computer Science
Location: Poster Session 1, 10:30 AM - 11:15 AM: Campus Center Auditorium [A3]

During their time in college, students often struggle to juggle various responsibilities, ranging from coursework and college and job applications to academic planning, extracurricular activities, and more. Academic advisors work together with students to manage these responsibilities; however, their knowledge bases do not fully overlap. It can be difficult to provide specialized advice when an advisor has many students to keep track of. In this project, we attempted to answer the following question: To what extent can we bridge this gap using an agentic AI workflow that emphasizes personal context to deliver specific, targeted advice? 

We designed an agentic workflow utilizing the LangGraph framework to orchestrate communication between LLM agents and provide them with access to useful tools. Through a web application, users can interact with chatbots and input personal details. The LLM interfaces with a SQLite database that stores important academic information, as well as chat logs to retain personal goals and contextual information. This information is shared between students and their advisors to enable effective communication. 

We expect this program to have a marked influence on both student and advisor workloads by making crucial information more readily accessible, allowing them to make decisions more confidently and efficiently. We conclude by discussing how AI systems can be used to enhance, rather than replace, meaningful human interaction.

RELATED ABSTRACTS

Agentic AI "Wayfinder"
Presenter: Edgar U. Falfan
Group Members: Michael Mahoney
Faculty Sponsor: Hao Loi
School: Quinsigamond Community College
Research Area: Computer Science
Location: Poster Session 1, 10:30 AM - 11:15 AM: Campus Center Auditorium [A4]

Traditional travel planning required users to manually coordinate transportation, accommodation, meals, and activities across multiple platforms—a time-consuming and often overwhelming process that continued throughout the trip. This project addressed the following question: How can autonomous AI help improve the pre-trip logistics under budget and preference constraints? The hypothesis was that an autonomous AI agent could optimize travel experiences by proactively managing routine tasks and providing real-time, contextually relevant recommendations.

The methodology involved developing a travel assistant system where users input their itinerary with dates, destinations, budget constraints, preferred attractions, and personal preferences regarding food and shopping. The AI agent operated autonomously throughout the trip, continuously comparing transportation options between destinations, identifying accommodations within budget, and discovering points of interest aligned with user preferences that travelers might not have initially considered. The system integrated historical and real-time weather data to recommend optimal daily activities, located nearby restaurants, convenience stores, and public restrooms based on the current location, and suggested contextually relevant shopping opportunities.

The results demonstrated that autonomous AI significantly improved travel efficiency by automating multifaceted tasks that typically required constant human attention. The system saved time through automated comparisons and recommendations and enhanced the overall travel experience by highlighting opportunities that travelers might otherwise have missed. This research contributed to understanding how autonomous AI agents can transform everyday experiences by managing the operational complexity of travel, allowing users to focus on enjoyment and exploration.

RELATED ABSTRACTS

Using Technology to Address Imposter Syndrome in Computer Science Students
Presenter: Dev Vinod Mehta
Faculty Sponsor: Ellen Correa
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 1, 10:30 AM - 11:15 AM: Campus Center Auditorium [A40]

Imposter syndrome is highly prevalent among students in computer science, where competitive academic cultures, “genius” narratives, and limited transparency around pathways to success can undermine students’ confidence and sense of belonging. This civic engagement project addresses imposter syndrome within the Manning College of Information and Computer Sciences (CICS) at the University of Massachusetts Amherst by developing a community-centered technological platform that connects students to relevant courses, research groups, student organizations, and career development opportunities. The problem motivating this work is not only individual self-doubt but also structural inequities that disproportionately affect students from underrepresented and first-generation backgrounds, limiting their access to information and community support.

 

The project employs an iterative, human-centered design methodology. A needs assessment will be conducted through anonymous surveys and focus groups to understand students’ experiences with imposter syndrome and barriers to engagement. Based on these findings, a prototype conversational tool or web-based platform will be developed, piloted, and refined through stakeholder feedback, including collaboration with the Public Interest Technology Initiative.

 

The guiding hypothesis is that increasing transparency and access to opportunities will strengthen students’ sense of belonging and reduce imposter-related self-doubt. In a broader context, this project demonstrates how community-informed technology design can address systemic barriers within higher education, advancing equity and inclusion by reshaping how students navigate and experience academic institutions.

Autonomous Photographer
Presenter: Carlos Vasquez
Faculty Sponsor: Komalpreet Kaur
School: Salem State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Concourse [B7]

In everyday situations such as travel, outdoor activities, or public events, capturing a professional photograph can be unexpectedly difficult. You are limited to taking a selfie with restricted composition, relying on a stranger passing by for assistance, or investing in a professional photographer. Each option presents compromises in quality, convenience, or cost. This project introduces an AI-driven robotic camera system designed to eliminate these trade-offs by automatically detecting subjects, optimizing composition, and capturing high-quality images independently.

The solution integrated a Raspberry Pi-based robotic system equipped with a camera module and pan-tilt servo motors. Computer vision techniques were implemented using OpenCV and lightweight deep learning models to perform real-time subject detection and tracking. A composition analysis module evaluated framing based on photographic principles such as the rule of thirds, symmetry, headroom, and exposure balance. A closed-loop control system translated the composition scores into motor adjustments, enabling the device to reposition itself automatically before triggering image capture.

The interface allows to initiate photo sessions and downloading stored images. Captured photos were evaluated using sharpness, exposure, motion blur, and contrast metrics, and metadata was stored locally in JSON format.

This project presents a functional prototype that combines embedded systems, computer vision, robotics control, and user interface design. 

RELATED ABSTRACTS

Viking P300: A P300 EEG Brain-Computer Interface for Real-Time Four-Command Game Control
Presenter: Roman Bukreev
Faculty Sponsor: Komalpreet Kaur
School: Salem State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Concourse [B8]

This project develops a P300 event-related potential brain-computer interface (BCI) that lets a player control a 2D maze game using EEG. In the game, the player navigates through a maze and collects items; movement is selected through a four-choice visual oddball paradigm. Directional arrow stimuli (Up, Down, Left, Right) flash in randomized sequences while the user focuses attention on the intended direction. EEG is recorded with a 16-channel g.tec g.Nautilus system across multiple data-collection sessions, producing labeled target and non-target flashes for supervised learning.

A Python pipeline segments EEG into stimulus-locked event-related potentials and extracts features for classification. Classification models such as regularized linear discriminant analysis and logistic regression are trained and evaluated using cross-validation strategies that test generalization across recording sessions. Performance is summarized with classification accuracy and command-selection speed.

The trained classifier is integrated into a real-time game loop to convert predicted targets into directional inputs with low latency, enabling hands-free maze navigation. The outcome is an end-to-end BCI game prototype and a reproducible analysis framework for comparing stimulus design choices and model performance. This work supports future improvements in accuracy, selection speed, and user comfort, and illustrates how ERP-based BCIs can be applied to interactive games that require reliable discrete command selection.

MataVision: Predicting and Preventing Traffic Accidents in Salem, Massachusetts Using Machine Learning
Presenter: FATOUMATA BARROW
Faculty Sponsor: Komalpreet Kaur
School: Salem State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Concourse [B9]

Over the past several years, Salem has experienced a noticeable rise in traffic incidents, especially during the fall when the city becomes packed with visitors. With the streets more crowded and people constantly moving between major attractions, certain areas consistently become more dangerous than others. The city has plenty of crash data, but there isn’t a clear way to analyze patterns, understand what triggers these incidents, or see how nearby points of interest contribute to higher risk.

The project focuses on using machine learning and geospatial tools to study the last five years of Salem’s crash records and identify where and when accidents are most likely to occur. By combining crash locations, POI proximity, and time-based factors, the aim is to highlight high-risk zones that the city could target for improvements. This is important because better insight into why crashes cluster in certain places can support safer street design, smarter traffic control, and more proactive planning, especially during busy seasons like October.

The project also includes exploratory data analysis, feature engineering, and predictive modeling to better understand patterns in crash severity and frequency. Visualizations such as hotspot maps and seasonal trend graphs will be used to clearly communicate findings to city planners and community stakeholders. By transforming raw crash data into actionable insights, this research aims to contribute to safer urban mobility and data-driven decision-making in Salem.

Java Uno Project
Presenter: Nason John Omasta
Group Members: Thomas Warren Willette
Faculty Sponsor: Ali Al-Faris
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C1]

This project is a Java text-based version of the card game UNO. Several custom classes were
crafted for this program, which include a Card class (which holds the attributes for UNO cards),
a PlayerHand class (containing a LinkedList of Card objects as well as insertion sorting
algorithms for ordering cards), a Player class (containing attributes such as the player’s name),
and a TurnOrder class (containing the active player order and the methods to switch both the
active player or reverse the order when a Reverse card is placed). Inside of the Main class,
there are various methods which set up two Stacks of cards for pulling and placing (and the
logic for both actions), as well as CPU logic, and victory conditions for the game.

RELATED ABSTRACTS

Crime Tracing Program
Presenter: Kaniz Fatima
Group Members: Anastasia Belle Hocurscak
Faculty Sponsor: Ali Al-Faris
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C2]

This project focuses on the design and implementation of a comprehensive Java-based crime data management and analytics system aimed at supporting law enforcement officers, legal professionals, and criminal investigators in making informed decisions. The system incorporates secure user authentication, utilizing a HashMap data structure to efficiently store and verify user credentials. This ensures fast access control while maintaining data security and integrity.

Crime records are imported from CSV files and processed through a dedicated Crime Data class, which organizes essential attributes such as crime type, location, date, time. By structuring the data into well-defined objects, the system enables efficient storage, retrieval, and manipulation of crime-related information.

The application generates detailed statistical reports to assist in crime analysis and strategic planning. These reports include crime rates categorized by location and type, neighborhood safety scores based on incident frequency, and time-based risk pattern analysis to identify high-risk hours or recurring trends.

To enhance coordination, Sets are used to maintain unique police department contacts, preventing duplication and ensuring accurate communication records. Additionally, the system supports dynamic data updates, allowing new crime records to be added or modified seamlessly. This ensures that the database remains current and reliable.

Overall, the system integrates data management, security, and analytics into a unified platform, improving investigative efficiency and facilitating better inter-agency collaboration.

Pet Adoption System Abstract
Presenter: Namra Waqar
Group Members: Milla Santos
Faculty Sponsor: Ali Al-Faris
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C3]

Our pet adoption system oversees and manages the pet adoption process using Java programming, such as, managing customer profiles, managing pet profiles, allowing customers to join the queue for adoption, and also giving customers recommendations on what pets they can adopt based on their preferences. While most pet adoption systems have the staff members managing customer profiles, our system has a mini self-service system for customers to create or edit their profile themselves, saving the staff members’ time and increasing the efficiency of the pet adoption process. Customers also have a choice of entering their preferences on what type of pet they’d like to adopt. They can choose a specific species, breed, and a personality trait. The system uses data structures to organize all the information and  increase efficiency to ensure optimal performance. Linked lists are used to store the profiles of pets and customers. Hash tables are used to store the staff members’ usernames and passwords, and a queue is used to manage the waiting list for adoption. Our pair system, which recommends pets to customers based on their preferences, uses a linear search to go through the customer’s preferences and the pet’s traits to see if there is a match.
Sports Car Racing Strategy Assistant with Reinforcement Learning
Presenter: Shayna Mullett
Faculty Sponsor: Ali Al-Faris
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C4]

This project presents a Python-based AI race strategy assistant developed for Formula 1, designed to optimize race outcomes for a single car–driver combination. The system employs reinforcement learning to model strategic decision-making across race conditions, using historical data from the 2022–2024 seasons as its training environment. By learning through iterative reward-based optimisation—where finishing position and race time serve as key performance signals—the agent develops strategies for tyre selection, pit stop timing, and stint management under varying competitive scenarios.

The reinforcement learning framework enables the agent to evaluate long-term trade-offs, such as track position versus tyre degradation, and to adapt its policy based on evolving race contexts. The model is trained on historical race states and outcomes, allowing it to approximate realistic race evolution while learning optimal responses to both predictable and stochastic events.

A central feature of the system is its dynamic adaptability. During simulation, users can manually introduce race events such as safety cars, punctures, lock-ups, or time penalties. These interventions modify the race state in real time, prompting the agent to recompute and adjust its strategy to maximise finishing position under updated constraints.

By combining historical data-driven modelling with reinforcement learning and interactive adaptability, this project demonstrates the potential of AI-assisted strategic decision systems in high-performance motorsport and other dynamic, uncertainty-rich environments.

FunGames.Java
Presenter: Hieu Minh Nguyen
Faculty Sponsor: Ali Al-Faris
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C5]

          The increasing interaction between social media and online gambling has raised concerns about its impact on youth behavior and culture. As social media platforms and the internet adapt to their users, major industries are using the internet to adapt to their consumers as well. Gambling has been widely regarded as controversial in the United States due to its harmful social and financial consequences. However, as the internet has become more accessible and widely used, gambling has gained its footing on the internet as well. Online gambling platforms have expanded significantly, with teens and young adults among the most targeted demographics. This study surveys students and adults, as online gambling platforms are accessible to users across age groups. The influence of gambling does not necessarily only connect with the casino or sports. Video games have increasingly incorporated addictive monetization strategies built by profit driven organizations. They can come in the form of a reward for spending a certain amount of money in the game or opening enough mystery boxes to obtain certain items. As part of this project, we developed a simulated casino game using Java, incorporating different data structures and a simple G.U.I for user experience. Our goal is to reduce the use of online gambling and to spread awareness on the addictive and harmful effects of gambling through simulated games.
Chernobyl Pet Simulator
Presenter: Alexa M. James
Group Members: Israa Touaiher
Faculty Sponsor: Ali Al-Faris
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C6]

In this game, players are attempting to keep their mutant pet alive by playing games, earning money, giving affection, buying clothes and food, and even treating their pet’s tumors. Players can also view their wallet, check high scores through the leaderboard, and view all of the items they have bought.


This project aimed to familiarize and adapt Java data structures like HashMaps, Stacks, and ArrayLists, as well as algorithms like searching and sorting, to create a digital pet with an easy and fun user interface. We also incorporated vibrant text boxes and ASCII art with JavaFX to make the UI more interactive and interesting.


To create this game, we had 5 classes, including the pet, shop, item, leaderboard, and game classes, as well as methods and constructors. These classes helped us stay organized and make the code more intuitive and easier to edit and maintain. For instance, our pet class has the feed and clothes methods, which allow players to buy food and clothes to add to their pet’s pantry and closet. The caress method allows players to show affection for their pet, which improves health and happiness. The check alive method checks that the health, happiness, and wallet stats are above 0. If any of these stats are 0, the pet will die, and the game will exit. The decay method lowers the health and happiness stats throughout the game. And finally, the showstats method displays the pet’s happiness, health, name, and wallet. 
Implementation of a Relational Database for Managerial Accounting Analysis
Presenter: Stephen Kihanya Nganga
Faculty Sponsor: Nada Al Sallami
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C17]

The aim of this research is to design and implement a relational database using a real-world managerial accounting dataset to support financial analysis. The dataset used in this project was obtained from Kaggle (Hassan, 2023) and contains realistic accounting information, including transactions, departments, categories, revenues, and expenses. The project began with cleaning and converting the original Excel file into CSV format to resolve data inconsistencies before importing the data into MySQL Workbench, demonstrating the importance of proper data preparation in accounting systems. A relational database schema was designed using entities such as Departments, Categories, Projects, Locations, Accounts, and Transactions, and an ER diagram was created to illustrate the relationships among these entities. SQL queries were developed to analyze expenses and calculate totals by department, and a stored procedure was implemented to retrieve department-specific transaction data.

The database was then connected to a Java application using JDBC, allowing the stored procedure to be executed from an external program. While challenges were encountered during the Java integration, particularly with connection setup and runtime input, the system successfully demonstrated how relational databases support managerial accounting analysis and decision-making. Future work could expand this system by incorporating additional datasets, advanced analytical queries, enhanced reporting features, and improved user interaction within the Java application to further support managerial decision-making.
Data Science Career Database
Presenter: Karol Harasim
Faculty Sponsor: Nada Al Sallami
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C18]

The aim of this research is to construct a relational database system that provides in-depth and precise analysis of individual job postings in the data science field. Understanding the trends in data science careers can prove to be tedious when there are several job related attributes. A CSV file can assist in analysis but is inefficient for making large scale comparison and analysis. Constructing a relational database based on an imported CSV file provides the structure and organization to make precise analyses and draw conclusions on data science careers. Relational databases are data structures that store normalized data into separate tables that enable an easy way to clean the data to reduce redundant, missing, and incorrect data. Originally starting with one table consisting of all of the data from the CSV file, processing the data using MySQL allowed for the creation of four tables including an associative table between the data jobs and skills tables. Giving each instance of a company, job, and skill a unique identifier decreased redundancy of companies and skills while each instance remained correlated to its respective job. Creating and executing queries on the newly created database quickly answers questions such as determining the most sought after skill in data science job postings. These queries grant data science job candidates the ability to analyze job market patterns. The database demonstrates how CSV file data can be transformed into a structured format to make meaningful observations in the data science job market.

RELATED ABSTRACTS

Bug Attraction to Light
Presenter: Arianna Mary Nevis Michael
Group Members: Cassandra Jreij
Faculty Sponsor: Nada Al Sallami
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C19]

The rise of LED lights as a means of outdoor lighting is typically seen as positive because it is cheap and requires less maintenance. However, LED lights produce a white or blue light that attracts insects. This artificial light disrupts their circadian rhythm and can lead to their death as they circle the light endlessly. The purpose of this project is to find which of these various factors (temperature, humidity, moon visibility, etc.)  would affect the attraction of insect orders to a light. Using a three table structure (Entity Relationship Diagrams), we defined the types of relationships between three mains sections of data. Our research utilizes an open source database and MySQL Workbench software to provide data modeling and create a network of tools (stored procedures, complex queries, etc.) for analyzing various factors influencing the attraction of bug orders to different light sources. We formatted these tools around the tables formed from the Entity Relationship Diagrams. Specialized operators, such as UNION ALL, are used to create a smoother return. We use these in conjunction with the stored operations to return very specific results. Our procedures result in an organized and easily navigable table for each of the specific queries, making the information from the bug database easier to analyze. Further work in this project would help analyze multiple factors at input and arrive at a more focused result. Overall, this data can help create a solution to prevent bug fatality.

RELATED ABSTRACTS

Risk-Aware Parking Lot Recommendation Using Greedy and Weighted Selection Algorithms in a Smart Campus Environment
Presenter: Rick Manuel Woubinwou Djouwe
Faculty Sponsor: Nada Al Sallami
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 163 [C20]

Efficient parking allocation remains a persistent challenge on university campuses, particularly during peak arrival periods when students must search repeatedly for available spaces. This study presents the design and evaluation of a risk-aware parking lot recommendation system modeled for a smart campus environment. The system dynamically simulates parking lot occupancy based on time-of-day arrival and departure patterns and recommends optimal lots according to a selected destination building.

Two allocation algorithms are implemented and compared: a greedy closest-distance approach and a weighted risk-aware selection algorithm. The greedy algorithm prioritizes minimal walking distance to the destination building, while the weighted algorithm incorporates both walking distance and a probabilistic estimate of lot saturation to reduce the likelihood of arriving at a full lot. Each algorithm ranks available parking lots using a cost function and returns the top recommendations in real time.

Experimental evaluation examines computational performance, average walking distance, and sensitivity to risk weighting under varying congestion conditions. Results demonstrate that while the greedy strategy minimizes walking distance, the risk-aware algorithm provides more balanced recommendations during peak hours by reducing exposure to high-occupancy lots.

This work illustrates how algorithmic decision strategies can improve campus parking efficiency and highlights the practical role of cost-based optimization in smart infrastructure systems.

Raspberry Pi Based Solar Power Monitoring System
Presenter: Philip Paskal
Faculty Sponsor: Manish Wadhwa
School: Salem State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 165 [D11]

Small-scale solar systems are becoming more popular for residential and off-grid applications, but many lack the intelligent monitoring capabilities that some users may want. Users will often depend on either standalone charge controllers that regulate charging but provide limited system visibility and minimal automation beyond what the manufacturer provides, or the systems are expensive and can lock users into vendor product catalogues. This project presents the design and implementation of a Raspberry Pi based solar power monitoring system. The Pi communicates with the charge controller to retrieve voltage, current, and battery status information. A Python script processes these reported values to calculate instantaneous power output and estimate battery charge percentage. Benchmarks for success include accurate interpretation of controller data, stable operation, and recovery after power interruption. By integrating existing off-the-shelf hardware, this project demonstrates a practical and scalable approach to enhancing the efficiency, safety, and intelligence of small-scale solar energy systems.

Pi In The Sky Birdhouse
Presenter: Diane M. Jubeili
Faculty Sponsor: Manish Wadhwa
School: Salem State University
Research Area: Computer Science
Location: Poster Session 2, 11:30 AM - 12:15 PM: Room 165 [D12]

This project presents an automated birdhouse monitoring system designed to provide real-time wildlife identification through the integration of embedded hardware and cloud-based artificial intelligence. Built on a Raspberry Pi platform, the system utilizes a PIR (Passive Infrared) motion sensor to detect the arrival of a bird. Upon triggering, a Pi Camera module captures high-resolution imagery, which is then transmitted to the Google Cloud Vision API.

By leveraging machine learning through this API, the system identifies the specific species or characteristics of the bird. The processed information, along with the captured image, is then delivered to the user via SMS notification. This end-to-end IoT pipeline demonstrates the practical application of sensor fusion, cloud computing, and automated remote communication, offering a sophisticated yet accessible solution for ornithological study and hobbyist wildlife observation.


College Success Factors
Presenter: Alivia Glynn
Group Members: Zachary Albert Kimball
Faculty Sponsor: Elena Braynova
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Campus Center Auditorium [A39]

Abstract: In this project we studied a College Placement dataset consisting of 10,000 student records focusing on the factors that determine students’ success such as academic performance, cognitive ability, experience, and placement outcomes. The dataset was analyzed using a variety of visualization, statistical analysis and Machine Learning techniques. Using R, we created histograms, boxplots, scatterplots, bar charts, and correlation heatmaps to explore patterns and relationships. Visualizations revealed a notably strong positive relationship between Previous Semester Result and CGPA. Internship experience appeared to relate to higher placement rates, while IQ and other cognitive measures showed relatively weak associations with academic outcomes. Statistical analysis supported these observations. The correlation between Previous Semester Result and CGPA was approximately 0.98, and linear regression produced an R2 value near 0.96, indicating that prior academic performance explains much of the variation in current GPA. Hypothesis testing showed no statistically significant difference in GPA based solely on internship experience. Using the WEKA tool we looked deeper into the dataset and the patterns there. Using the Decision Tree Classification methods we achieved very high accuracy in predicting Academic Performance, while regression models effectively estimated Projects Completed and Previous Semester Result. In contrast, predicting IQ resulted in higher error rates, suggesting that cognitive measures are not easily inferred from academic or experiential experiences attributes. Overall, the findings highlight the strong role of academic consistency in student outcomes and provide insight into the relative influence of experiential and cognitive factors.

RELATED ABSTRACTS

Lifestyle, Health, and Sleep Quality
Presenter: Sean Wang
Faculty Sponsor: Elena Braynova
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Campus Center Auditorium [A40]

In this project we analyzed sleep quality, health, and lifestyle data to explore relationships between sleep habits, demographics, lifestyle, and a variety of health indicators. After preprocessing, the data was analyzed using a wide variety of visualization, statistical analysis and Machine Learning techniques. We discovered strong correlations between certain attributes, such as sleep duration and sleep quality, sleep quality and stress level, and more. We looked at the data deeper using Machine Learning Classification, Association Rules and Numerical Prediction methods. We found that with very high accuracy we could predict a participant having a sleeping disorder or not. Using a variety of Classification models and methods we found out that the female and male groups in the dataset have different sleeping patterns. Association Rules mining revealed some interesting relationships between gender, sleeping habits, health and lifestyle indicators. Our results on predicting sleep quality using Numerical Prediction methods are also interesting. They show how the sleep quality depends on sleep pattern and health attributes. 
Experience Report: Software Development Teams with Human-AI Collaboration
Presenter: Adrianna Frazier
Faculty Sponsor: Karl R. Wurst
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C1]

Kits are educational tools built to provide an authentic environment to learn Computer Science concepts. Kits consist of an open source project with student activities and instructor guides. The MicroservicesKit allows students to work with and understand modern complex program integration and containerization while the GitKit provides students with experience using Git, the dominant version control system. We are currently refactoring, improving, and generalizing the tool used to build and deploy the kits. To do this, our team is coming up with a top-level design that will decompose GitKits’ monolithic architecture. The goal is an architecture where each task is a single, atomic-function building block that together form a larger cohesive build. With this reconstruction, we aim to have the system move from working only with GitHub to one that also has GitLab integration. By using Agentic AI to help develop our code, it has helped us develop the individual components, create a plan, implement changes, and test to make sure everything is integrated correctly and works as desired. In conversation with the AI tool we can brainstorm, design, plan implementation, write and run tests, and implement the code. We can do this in a fraction of the time, with higher quality, than it would take a human team member. As we continue to use the AI tool, we are learning how to make it work better in this project and in future projects.
Evaluating Algorithmic Fairness in Credit-Scoring Systems: A Machine Learning Analysis of Socioeconomic Bias in Financial Technology
Presenter: Arianna Alexandra Fernandez
Faculty Sponsor: Daniel Haehn
School: UMass Boston
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C2]

As machine learning systems increasingly shape financial decision-making, concerns have emerged regarding algorithmic bias in credit-scoring models. This study investigates whether predictive models for credit default risk disproportionately affect demographic and socioeconomic groups, potentially leading to unfair credit denials. To address this, a fairness-aware machine learning framework will be developed, considering structural relationships among variables.

 The study analyzes different algorithms and datasets, including the credit models used by the Board of Governors of the Federal Reserve System, the Default of Credit Card Clients dataset from the UCI Machine Learning Repository, and Kaggle datasets such as Home Credit Default Risk and Give Me Some Credit, utilized by Home Credit to promote financial inclusion, among others. It compares performance across these datasets and conducts thematic analyses to identify variables linked to unfair algorithms. By examining model coefficients and feature importance measures, this research assesses the extent to which algorithmic credit assessment may reinforce existing inequalities. The study explores how the linear independence of variables, such as demographics, can significantly influence risk percentage calculations and examines how certain variables, such as peer-based features that tie default risks to the profiles of neighboring individuals, can promote financial exclusion. 

After completing the analysis, a machine learning model will be developed to illustrate how a fairer model would perform, comparing its results with those of existing models to evaluate efficiency. Finally, a market analysis will be conducted to assess the long-term effects of implementing the new model, aiming to demonstrate how a fairer model would benefit society through market inclusion, thereby fostering the economy.
Designing AI-Assisted and HCI-Driven Learning Solutions to Support Home-Based Learning for Students with Multiple Sensory Impairments (Blind and Mute) in Sub-Saharan Africa
Presenter: Serin Kitery
Faculty Sponsor: Amanda Potasznik
School: UMass Boston
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C3]

Assistive technologies are products that enable individuals’ health, inclusion, and participation in society through the improvement of their functionality, including mobility, communication, vision, and hearing. In developed countries like the United States, substantial study and research of and with disabled populations has resulted in the strong development of assistive technologies. We see impressive design principles and advanced AI-assisted software on technologies for children with sensory disabilities in education, including tools like Described and Captioned Media Program(DCMP) online notetaker training module for deaf students, Braille displays and notetaker for visually impaired students, and text-to-speech software for mute students with personalized learning. However, in emerging markets like African countries, students with disabilities still face challenges in education. There are nearly 29 million children, mostly with physical and sensory disabilities, in Eastern and Southern Africa alone, and less than 10% of them attend schools. Those who do manage to attend school perform poorly on average in subjects like Mathematics and English that require reading. The demand is clearly high for improved inclusive education in African countries. This study aims to propose human-centered paradigms for AI-assisted tools that support home-based learning for students with visual and speech impairments in Sub-SaharanAfrican countries. This study explores interactive Human Computer Interaction(HCI) design principles, offline-first AI models, and multimodal interaction techniques that can allow education inclusivity and culturally appropriate assistive technologies for students in environments with limited infrastructure. 
Breast Cancer Analysis via Machine Learning of Medical Images
Presenter: Elio Ngjelo
Faculty Sponsor: Maxim O. Lavrentovich
School: Worcester State University
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C4]


Predicting breast cancer recurrence remains a critical challenge in clinical oncology. We present a machine learning framework designed to predict disease progression using histopathological whole-slide imaging. We collect a dataset of 444 images from 47 patients, with each patient contributing 10 Region of Interest (ROI) images. The ROI images are slices through the cancerous tissue at a fixed resolution and contain both cancerous and regular cells. A standard Support Vector Machine (SVM) model, trained on unedited images, serves as our baseline model. We then discuss biophysical features, such as alignment of extracellular fibers and cell density, which may be used to train enhanced machine learning models.  By comparing the enhanced models with the baseline, we may identify which morphological patterns are most predictive of cancer recurrence.
Emotion Detection from Speech Audio Using Deep Learning Architectures
Presenter: Tanmay Sonawane
Faculty Sponsor: Alfa Heryudono
School: UMass Dartmouth
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C5]

Human speech carries rich paralinguistic information, particularly emotion, which provides valuable insight into psychological state, intent, and behavioral response. This project investigates how modern deep learning architectures can detect emotion from speech audio using time-frequency representations stored as numerical arrays. Centered on the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS), the methodology involves systematic preprocessing of raw audio files into normalized Mel-based NumPy (.npy) representations, followed by multimodal learning that jointly processes raw waveforms and spectrogram arrays. A pretrained ResNet-18 architecture is employed as the primary convolutional backbone, implemented within the PyTorch framework. Model performance is evaluated using accuracy, F1-score, precision, recall, and confusion matrices, with an achieved baseline accuracy of approximately 70% on a five-class emotion mapping.

To further assess robustness and generalization, this work will be extended to the Toronto Emotional Speech Set (TESS), enabling cross-dataset evaluation and combined-training strategies. In addition to PyTorch-based models, equivalent architectures will be implemented and tested using TensorFlow to provide a comparative analysis of deep learning frameworks for speech emotion recognition. Differences in training dynamics, performance, and deployment considerations across frameworks will be systematically examined. The ultimate application focus of this research is emergency service call analysis, where real-time emotion detection can assist dispatchers by identifying heightened stress, fear, or distress in callers. By benchmarking models across datasets and frameworks, this project aims to support the development of reliable, emotion-aware systems for safety-critical, interactive, and assistive technologies.


RELATED ABSTRACTS

Robert, a Nintendo R.O.B. simulator
Presenter: Lucas Newton
Faculty Sponsor: Allan Brockenbrough
School: Salem State University
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C6]

In 1985, the Nintendo Entertainment System (NES) game console was released in the United States. Along with the system itself, it was available in a bundle with a toy robot called R.O.B., the Robotic Operating Buddy. The purpose of R.O.B. was to help you play two specific games: Gyromite and Stack-Up. In each of these games, R.O.B. would look at the TV for a specific sequence of flashes, which would then tell the robot to act in a specific way. Nowadays, it’s impossible to play the games which required R.O.B. properly without having an actual unit, which have become hard to find and expensive.

Robert is a software-based simulation for R.O.B. which aims to solve this problem. The main part of Robert is an app made in the Godot game engine which features a 3D model of R.O.B. that can move around and act in an identical manner to the real unit. Because R.O.B. needs to communicate with a NES console to do anything, Robert supports two ways of connecting to a running system: either to a software NES emulator through use of a special plugin, or to a real NES console with a microcontroller.

The purpose of Robert is to be a proper historical preservation of R.O.B.. Potential users could range from museums which have the program running as part of an exhibit of historical video games, to home users which never had the opportunity to use R.O.B. in the first place trying it for the first time.


RELATED ABSTRACTS

Scaling Laws in Small Convolutional Neural Networks: Width, Depth, and Dataset Size Under Fixed Compute
Presenter: Joshua River Johnston
Faculty Sponsor: Ali Azhari
School: Bunker Hill Community College
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C7]

In machine learning, neural scaling laws empirically describe predictable power-law relationships between model size, dataset size, compute, and the network's overall test loss. There has been extensive testing done on large language model and transformer models with billions of parameters, but it is unclear if these laws, being asymptotic, can also be used to describe smaller convolutional networks. This work investigates scaling behaviors of small CNNs (between 18k-4m parameters) trained on various portion of the CIFAR-10 dataset. 
Model width, depth, and dataset size were systematically varied, and test loss was analyzed using power-law regression. Strong scaling relationships were observed between loss and parameter count (R² up to 0.996), with width scaling exponent α ≈ 0.16 and depth scaling exponent α ≈ 0.26. Dataset scaling exhibited a larger exponent β ≈ 0.43 (R² ≈ 0.99), indicating greater sensitivity to data size than model size.
The strength of these scaling relationships enables evaluation of model–data allocation under fixed compute constraints. Using estimated values of α and β, we derive predicted optimal ratios for parameter count and dataset size under a fixed training budget and empirically test these predictions.
If validated, this approach would provide a practical framework for compute-efficient training, allowing researchers with limited resources or large experimental workloads to reduce reliance on exhaustive hyperparameter sweeps.
  






RELATED ABSTRACTS

A Modular Architecture for Adaptive and Personalized Cognitive Prosthetics
Presenter: Candace Williams
Faculty Sponsor: Poonam Kumari
School: Bridgewater State University
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C8]

Personalization of assistive technologies is critical for adaptive cognitive support. However, current tools often rely on systems that lack closed-loop personalization and use strategies that focus on the characteristics of a general population, not each individual person. One method for the system to adjust itself to each person is by using neural signatures in tandem with deep learning, something that can result in a unique system for each human being. The primary goal of this research is to create a modular closed-loop architecture for personalized cognitive augmentation and rehabilitation using multimodal physiological signals and latent state modeling. This architecture combines electroencephalogram and electrocardiogram signals into a vector which represents the cognitive state of the individual at a given point in time. Knowledge distillation is then used to create a smaller model that can be deployed on a wearable device. The representation of cognitive state is used to create a reinforcement learning policy that an adaptive mechanism uses to adjust suggested health interventions. This is formally specified as a pipeline encompassing the following modules: fusion, distillation, calibration, and adaptation, with mathematical formalization and data flows. Thus, the proposed architecture will provide a personalization approach using latent cognitive state as a control variable and a deployable wearable strategy.

RELATED ABSTRACTS

Inferring Regulatory Mechanisms from Differential Gene Expression
Presenter: Ava Sajjadi
Faculty Sponsor: Kourosh Zarringhalam
School: UMass Boston
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C9]

Understanding pathological states requires moving beyond differential gene expression to identify upstream regulatory mechanisms and disrupted signaling pathways. Although conventional transcriptomic analyses detect genes that are significantly up- or down-regulated, they do not directly reveal the transcription factors (TFs) and regulatory interactions responsible for these alterations.

This project presents a unified computational platform for gene regulatory network inference that integrates two complementary statistical paradigms: a frequentist framework (QuaternaryProd) and a Bayesian network model (NLBayes). The system accepts transcriptomic signatures, harmonizes gene identifiers, and evaluates regulatory consistency against curated interaction networks. By incorporating regulatory directionality, network structure, and statistical uncertainty, the platform infers candidate TF activities and perturbed pathways associated with disease states.

Implemented as a researcher-facing web application, the framework standardizes outputs across both models and enables flexible comparison, filtering, ranking, and visualization of inferred regulatory interactions. This design allows users to examine how different statistical assumptions influence mechanistic interpretation while maintaining methodological rigor and transparency. By shifting analysis from gene-level changes to regulator-level inference, the platform enhances interpretability of transcriptomic data, supports hypothesis generation, facilitates identification of candidate therapeutic targets, and provides a scalable and reproducible computational foundation for systematic investigation of disease-associated regulatory mechanisms across diverse biological contexts.

RELATED ABSTRACTS

Prophet Inequalities in the Non-IID and Free-Order Setting
Presenter: Om Rakesh Mehta
Faculty Sponsor: Hedyeh Beyhaghi
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C10]

How well can an online decision-maker perform compared to an all-knowing prophet? This question sits at the heart of prophet inequalities, a framework from optimal stopping theory that studies decision-making under uncertainty. A classical result guarantees that when information is revealed sequentially, a careful player can always secure at least half the expected reward of the benchmark - a surprisingly strong guarantee given the uncertainty involved. Stronger guarantees are known when future outcomes are drawn from identical distributions, but this assumption is often unrealistic in real-world applications such as online advertising auctions and hiring.

This project investigates how performance guarantees change when outcomes are drawn from non-identical distributions. We first study the simplest non-trivial case involving two distinct distributions, where the decision-maker may also choose the order in which outcomes are observed. Our methodology reduces continuous distributions to carefully chosen discrete cases and analyzes threshold-based decision strategies, enabling tractable worst-case analysis. Our main result shows that in the two-distribution setting, the worst-case performance guarantee improves to 0.8 - which exceeds the classical bound of 0.5 and falls short of the best-known guarantee in the identical-distribution case.

This suggests that distributional heterogeneity imposes a measurable cost on online decision-making. We are currently extending these techniques to three or more distributions, guided by the hypothesis that non-identically distributed guarantees are strictly worse than those under identical distributions. More broadly, this work helps bridge the gap between idealized theoretical models and the uneven conditions encountered in practical decision-making systems.
Understanding Adaptation in Sequential Decision Learning Under Non-Stationarity
Presenter: Devina Gera
Faculty Sponsor: Greg Gregory
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C11]

Sequential decision problems arise in many real world domains where models must act on evolving, noisy data while accounting for feedback between predictions and actions. These settings introduce challenges including distribution shift, partial observability, and complex adaptation dynamics that are not fully captured by static predictive modeling. This thesis investigates how machine learning systems behave in non stationary sequential decision environments, with emphasis on stability, generalization across regimes, and sensitivity to uncertainty in model outputs. Financial markets serve as a representative high noise testbed in which signals are heterogeneous, environments evolve over time, and decisions influence future observations. The project explores how different learning approaches, including sequence modeling, probabilistic inference, and decision focused learning, affect downstream behavior in such environments. Rather than optimizing domain specific performance, the work focuses on understanding the learning dynamics that govern robustness, adaptation, and consistency of decisions over time. Through empirical analysis across simulated and real time-series settings, the thesis aims to identify methodological principles that support reliable decision making under shifting conditions. The broader goal is to contribute insight into how machine learning systems can be designed to operate effectively in complex sequential environments characterized by uncertainty and change.
Measuring and Evaluating AI-Generated Content on TikTok
Presenter: Jennifer Ye
Group Members: Manya Mehta, Tanish Gupta, Arshiya Sharma
Faculty Sponsor: Ethan Zuckerman
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C13]

As generative video technologies become increasingly accessible, AI-generated media is rapidly proliferating on social platforms such as TikTok, raising questions about platform disclosure policies  and the reliability of their automated labeling systems. This study examines the prevalence of AI-generated videos on TikTok and evaluates the performance of the platform’s AI-labeling mechanisms. Using a random sample of TikTok videos collected before and after the emergence of AI-generated video, we developed a technical pipeline that extracts video metadata, creator annotations, and engagement metrics from hundreds of videos. Rather than treating any single indicator as ground truth, we compare multiple signals of AI generation including creator-provided tags, platform-applied AI labels, and classifications produced by a high-accuracy independent visual detection model. By examining the agreement and divergence among these signals, we assess how reliably AI-generated content is identified on the platform. Using the resulting dataset, we analyze engagement outcomes, including views, likes, shares, and comments, to explore whether AI-related signals are associated with different patterns of visibility and interaction. The study provides an empirical evaluation of TikTok’s AI disclosure ecosystem and contributes evidence to broader discussions of accountability, trust, and governance in rapidly evolving social media systems.
Risk-Averse Stochastic Package Queries to Build Stock Portfolios
Presenter: Sanya Thakur
Faculty Sponsor: Peter Haas
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C14]

Given the wealth of investment options available, and the many potential outcomes associated with each purchase, stock market investing can be daunting for everyone, especially for beginners and for communities historically underrepresented in finance who may have limited access to formal financial education. Stochastic Package Queries (SPQs) offer a user-friendly and efficient solution, allowing everyone to simply express their decision-making problems, and receive optimal recommendations under uncertain conditions in seconds. Under the hood, SPQs are processed by a highly scalable system comprising stock databases, linear program optimizers, market simulators, and customized algorithms that work in tandem to derive personalized and optimal investment strategies based on millions of simulated possible worlds. Unfortunately, SPQs can only optimize expected returns, and thus, when applied to portfolio construction, often recommend risky investments that look strong “on average”, but perform poorly in bear markets due to limited diversification.

We enrich the class of problems that can be solved scalably using SPQs, by allowing them to return risk-optimized portfolios that minimize expected losses in their worst case outcomes, or the Conditional Value-at-Risk (CVaR). We derive a mathematical formulation of the CVaR optimization problem based on numerous simulated possible worlds of future returns. Using auxiliary variables, our formulation represents tail-loss exceedances in a linear, optimization-friendly form. This enables CVaR to be optimized within the same Integer Linear Programming structure used by SPQs, preserving the original feasibility constraints and search process while explicitly prioritizing protection against extreme downside outcomes.
Sampling Methods for Sublinear Low-Rank Approximation
Presenter: Aaron Zhengyuan Tian
Faculty Sponsor: Cameron Musco
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C15]

In modern data analysis, it is common to encounter matrices so large that we cannot afford to compute or store them in full. A canonical example comes from kernel methods in machine learning, where a dataset of n points induces an n x n kernel matrix measuring pairwise similarity between the data points. When the size of the dataset is large, forming all n2 entries is infeasible. Low-rank approximation addresses this bottleneck by replacing the matrix by a smaller, compressed form, leading to significant speedups in practice. For certain classes of matrices, including positive semidefinite (psd) matrices, it is possible to compute a high quality low-rank approximation with only a small number of matrix entries and in sublinear time.

We study sampling methods for sublinear low-rank approximation of psd matrices. Prior work has largely split into two camps: methods that are simple and fast in practice but come with weak error guarantees, and methods with strong error guarantees that are substantially more complex to implement. Recently, Randomly Pivoted Cholesky (RPCholesky) broke this barrier by producing a simple, empirically fast algorithm with strong trace error guarantees.

Motivated by this result, we ask whether an analogous approach is possible to attain Frobenius error guarantees. We give an empirical and theoretical analysis of randomized sampling methods for sublinear low-rank approximation. Our preliminary results show that Nyström methods, including RPCholesky, cannot achieve a (1+ε) relative-error guarantee in the Frobenius norm with few matrix entries. However, we hypothesize that Frobenius relative-error guarantees remain attainable via sampling according to approximate column norms.


AI-Generated Interpretive Feedback vs. Traditional Visualizations in Health Tracking: Effects on Perceived Engagement and Insight
Presenter: Anish Kamath
Group Members: Dedeepya Pidaparthi, Grace Zhou
Faculty Sponsor: Ravi Karkar
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C16]

The widespread adoption of smartphones and wearable devices has enabled users to collect extensive data on sleep, mood, diet, and physical activity. However, many health-tracking applications present data in raw or overly complex formats that increase cognitive load, limit interpretability, and contribute to user disengagement. This project investigates whether AI-generated interpretive feedback enhances perceived engagement, usability, and insightfulness compared to traditional visualization-based feedback in a health-tracking application designed for college students. We will develop two interactive UI prototypes grounded in HCI research, personal informatics theory, and mHealth literature: (1) an AI-enhanced version incorporating features such as LLM-based summaries, predictive insights, and conversational feedback, and (2) a non-AI version relying on standard charts and graphical data representations. Using a mixed-methods evaluation approach, 8–10 participants per prototype will complete structured usability tasks including logging health data, interpreting trends and predictions, identifying cross-domain relationships (e.g., sleep and mood), and navigating behavior-change recommendations. Think-aloud protocols, observational notes, and post-task interviews will capture confusion, motivation, skepticism, and perceived clarity. Participants will also complete the System Usability Scale (SUS) and Net Promoter Score (NPS) surveys. By integrating quantitative usability metrics with qualitative insights, this study aims to determine whether AI-generated interpretive feedback meaningfully improves user confidence, understanding, and motivation to consistently track and improve their health behaviors. 

 

SuDocu+: Refining Subjective Summarization Intent Through Natural Language Conversation
Presenter: Aryan Deshpande
Faculty Sponsor: Alexandra Meliou
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C17]

Subjective extractive summarization is critical for systems that help users digest large corpora—news, reports, and research, where what “matters” is refined over time. SuDocu is an existing subjective summarization system that captures the user's summarization intent as a constrained optimization problem (integer linear program) using user-created example summaries. The problem is then solved to produce the final summary with the optimal set of sentences. This approach is feasible for just one summary generation, but in reality, users refine their intent iteratively until they are satisfied. Requiring a new set of examples for every refinement is slow, cognitively heavy, and assumes the user already understands the latent topic structure of the corpus. Our goal is to extend SuDocu to support conversational, multi-turn intent refinement, enabling interactive, exploratory “what-if” summaries without requiring new example clicks.

We introduce an extension to the existing SuDocu infrastructure that we call SuDocu+, which will enable conversational capabilities for the user, through a lightweight natural-language refinement channel mapping a user's conversational commands (e.g., “more nature, less politics”, “emphasize recent developments”) into small auditable updates of package-query bounds and objective weights, explaining changes and flagging infeasibility if needed. In addition, we will contribute to warm-started, progressive optimization that reuses the previous solution, delivering fast updates with improvement guarantees and full provenance. At the center of SuDocu+ will be an intent-to-constraint mapping module that translates natural-language refinement into concrete adjustments to the underlying package query. This preserves the prescriptive guarantees of package-query optimization while enabling low-effort, iterative refinement in natural language.
Exploring User Demand and Design Requirements for a Weightlifting-Focused Wearable
Presenter: Jason Huang Alexander
Group Members: Ava Sokolosky
Faculty Sponsor: Madeline Lee Endres
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C18]

Commercial fitness wearables primarily emphasize cardiovascular metrics and general activity tracking, offering limited automated support for structured, set-based resistance training. While prior research has advanced repetition detection and intensity classification using wearable sensors, the real-world demand and design preferences of experienced young adult lifters remain underexplored.

This study investigates user demand, feature preferences, and pricing tolerance for a wearable designed specifically for weightlifting. We conducted an exploratory survey of 32 college-aged adults (mean age = 20.8) who work out an average of 4.38 days per week; 91% identified as intermediate or advanced weightlifters. While 59% reported not currently using a wearable during workouts, 87.6% expressed interest in a lifting-focused device. Frequently requested features included automatic rep detection, intensity classification, heart rate monitoring, set and rest interval tracking, and long-term performance analytics. Median willingness to pay was $100, with a notable subset indicating willingness to pay $200-300, suggesting both budget and premium market segments.

To contextualize these findings, we conducted follow-up semi-structured interviews with 10 participants (5 current wearable users, 5 non-users). This will highlight practical friction points with existing devices and guide the development of interview-informed user personas.

Combined quantitative and qualitative analyses informed two user-centered design concepts: (1) a streamlined ~$100 device tracking core weightlifting features, and (2) a premium $200-300 model offering enhanced sensing and analytics. Our preliminary findings suggest unmet demand for resistance-training-specific wearable technology and provide user-centered design insights for future wearable computing systems.
Alternative Packages through Constraint Perturbation
Presenter: Henry Philip Booth
Faculty Sponsor: Alexandra Meliou
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C19]

In the relational database setting, a package query returns a multiset of tuples that maximizes or minimizes a linear objective function subject to local and global linear constraints. Prior advancements have shown the correspondence of package queries to Integer Linear Programs(ILPs) and constructed algorithms like ProgressiveShading for reliable package query evaluation in billion-tuple datasets. However, existing package query algorithms treat constraint bounds as absolute thresholds, never returning packages that violate any constraint, even by a small margin. In practice, users may not be as particular about the constraint thresholds as the solvers, and would happily explore alternative packages that (i) significantly improve objective values at the cost of minor constraint violations, or (ii) significantly ease demand on resources through tightened constraints, with a small loss of objective value. In this paper, we propose an algorithm for alternative package recommendation through constraint perturbation. By formulating the alternative packages problem as a mixed-ILP, with individual perturbation variables for each constraint and total constraint perturbation bound, our algorithm can identify desirable alternative package recommendations that optimally balance package solution objective value with total constraint perturbation. We believe that our mixed ILP approach will outperform naive modifications of progressive shading to solve the alternative package finding problem, and are currently running experiments to confirm that hypothesis. Further, our algorithm recommends a selection of optimized and varied alternatives, providing users the flexibility to select the package that is most desirable to them.
Computing Tree Decompositions of Road Networks
Presenter: Patrick Martin
Group Members: Ariana Grace Guzman, FNU Gayatri, Rashi Jain
Faculty Sponsor: Hung Le
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C20]

Obtaining the shortest path between two vertices is a fundamental problem in graph theory and computer science. It is critical to modern navigation systems, transportation planning, and large-scale geographic information systems. Since classical methods such as Dijkstra’s algorithm are prohibitively slow on large graphs, indexing techniques are often used to increase the speed of shortest-path queries. Several modern indexing methods use tree decomposition (TD), an algorithm used to represent a graph as a tree. TD iteratively groups vertices of the original graph and then assigns adjacency between groups such that a tree is formed. The size of a TD’s smallest group minus one is its treewidth. Indices can then be constructed to solve shortest-path queries in O(treewidth) time. 

Our work focuses on large, sparse, undirected graphs known as road networks and builds upon iterative, elimination-based approaches for constructing TDs. Because computing a TD of minimal treewidth is computationally infeasible for large graphs, they are often computed using heuristics. There is currently a lack of data surrounding heuristic performance on road networks, and research surrounding TD heuristics often omits implementation details that can affect performance in real-world applications. We plan to evaluate the treewidth, treeheight, query processing time, and other metrics to compare several heuristic methods. In doing so, we aim to improve our understanding of which heuristics are optimal for producing TDs and indexes on road networks. 
Automated Wing Landmark Detection for Morphological Classification of Winter Moth Species
Presenter: Thuyen Pham
Faculty Sponsor: Jeremy Catalin Andersen
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C21]

In New England, the native Bruce spanworm, Operophtera bruceata, and the invasive Winter Moth, Operophtera brumata, coexist, share similar morphology, and hybridize extensively, thus complicating efforts to assess population dynamics and manage the invasive species’ spread. Traditional identification methods, such as genitalia morphology and DNA-based phylogenetics, are accurate but slow, labor-intensive, and unsuitable for large-scale monitoring. Preliminary evidence shows that wing venation landmarks differ between the two species, suggesting feasible automatic species-classifier based on wing venation. Previous works in vein detection use U-Net and CNN architectures for semantic segmentation, but these models demand dense annotations and intensive multi-layer computation during training and inference. In this project, we investigate a low-annotation alternative using a one-stage detector You Look Only Once (YOLO) on bleached wing images to (i) localize wings and (ii) support landmark-based feature extraction for classification. We train on 2000 images (1000 images per species) and evaluate using stratified split by species. YOLO performance is summarized using mAP@0.5, mAP@0.5:0.95, and F1-score. Ongoing work will compare the result from YOLO model investigation with the result obtained from state-of-the-art U-Net and CNN. We anticipate YOLO will provide the best tradeoff between accuracy and throughput for localizing wings, enabling rapid processing of large image batches. A segmentation model such as U-Net may yield more detailed wing boundaries, but at higher annotation cost and computational complexity. 
Finding the Top-l Arms with an m-Constrained Quantum Oracle
Presenter: Avni Gunjikar
Faculty Sponsor: Mohammad Hajiesmaili
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C22]

Quantum best arm identification (QBAI) studies how to efficiently find the best among a set of arms, each associated with a quantum oracle. Applications could extend to finding the best paths in quantum networks, based on different parameters such as fidelity and noise, and the extension of this to the top-l arms would allow for alignment with more realistic hardware etc. Specifically, we are attempting to find top l arms with the best means with the fewest queries, called query complexity. We will primarily be working with the m-constrained oracle, which has the ability to query m arms at once, achieving a greater speedup than both the classical oracle and our first quantum oracle.  We will be adapting and extending the existing solution for the m-constrained oracle which finds the best arm, to find the top-l arms. We will attempt to prove that this retains the same query complexity as the m-constrained quantum oracle up to a constant factor of l. This entails extending the existing algorithm to work for the top l arms instead of just the one, and proving that all subroutines in the existing routine either work for the top l arms or can be extended to work for the top l arms. Once addressed, our method would play a key role in improving the efficiency of future quantum systems, e.g., quantum networking.


Fair or Fail? Formal Methods for Evaluating ML Fairness
Presenter: Shriya Gautam
Faculty Sponsor: Heather Conboy
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C23]

As machine learning and software has become increasingly integrated in the lives of the average person, ensuring that this kind of algorithmic decision-making is unbiased and fair to all groups is no longer just a technical preference, but also an ethical necessity. The potential ramifications of using unfair or biased algorithms in real world use cases can be disastrous, which is why having a formal system for verification and evaluation is more important than ever. This study aims to explore the landscape of various Machine Learning fairness frameworks and constraints as they apply to various software or Machine Learning examples, with a focus on ensuring equitable outcomes across diverse demographics. Through the planned activities, this study will produce a comprehensive overview of the methods by which fairness can be measured and guaranteed.

To accomplish this, this study will review several existing fairness evaluation frameworks such as the Seldonian framework, and explore other measures that are currently used to track fairness and bias across demographic groups. The study will then culminate in the replication of key algorithmic experiments, and the application of these fairness-aware methods to new software use cases. The resulting overview will provide a roadmap for measuring, auditing, and enforcing fairness in modern software systems.


RELATED ABSTRACTS

Enabling Communication in Multi-Agent VLAs
Presenter: Dorian Benhamou Goldfajn
Faculty Sponsor: Shlomo Zilberstein
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C24]

Vision–language–action models (VLAs) have emerged as a promising approach for enabling general-purpose robots by jointly learning visual perception, natural-language understanding, and motor control within a single framework. However, training data requirements increase exponentially as action space grows, exposing a fundamental bottleneck in their ability to execute high-dimensional actions. Additionally, many real-world domains, including search-and-rescue and disaster response, inherently require multiple robots to coordinate actions and share information in real time. Despite recent progress, existing VLA architectures lack explicit mechanisms for inter-agent communication and coordination. 


This limitation motivates the development of VLAs that can communicate in multi-agent scenarios to distribute action complexity and extend overall system capability. This proposal aims to develop novel VLA architectures that support explicit communication among multiple agents, enabling real-time coordination and information sharing.  


Natural language provides an intuitive starting point for inter-agent communication, leveraging the language-model backbone within VLA architectures. However, long-form language is inefficient for high-throughput, low-latency robotic coordination. More effective alternatives include constrained vocabularies, structured symbolic messages, or compact latent representations that encode agent intent or planned actions, preserving semantic structure while enabling scalable real-time coordination. The proposed communication pipeline consists of annotating offline trajectories with a compact, predefined vocabulary or learning communication protocols from environment interaction through multi-agent reinforcement learning. The resulting models will be evaluated on a suite of scalable multi-agent robotics tasks developed in IsaacLab, designed to test coordination efficiency, generalization, and robustness across increasing task complexity.

RELATED ABSTRACTS

Intermittent Machine Learning for Battery-Free Environmental Sensing Under Solar Power
Presenter: Linh Dang
Group Members: Duc Ngoc Nghiem
Faculty Sponsor: VP Nguyen
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C25]

Battery-free sensing platforms powered by ambient energy harvesting offer a promising approach for long-term environmental monitoring without battery replacement or maintenance. However, these systems operate under highly unstable energy conditions, as available power varies with environmental factors such as lighting. This behavior conflicts with traditional machine learning assumptions that computation can run continuously. In this project, we investigate how lightweight on-device inference behaves when executed under intermittent power.

We developed a solar-powered prototype that activates only when its storage voltage exceeds an operating threshold. The device periodically collects environmental data and attempts local inference whenever energy is available. Because harvested energy fluctuates, the system repeatedly shuts down and restarts, leading to partial execution and inconsistent prediction behavior. A mobile logging interface was used to monitor voltage levels, execution cycles, and inference outcomes across repeated charge–discharge periods.

Initial observations show that sensing and inference perform reliably during stable illumination but degrade under dynamic lighting conditions. Frequent cold starts and incomplete computations were observed when energy availability was marginal, and prediction consistency decreased during these unstable periods. These results suggest that directly applying always-on inference strategies to energy-harvesting devices can lead to unreliable behavior.

This work highlights the importance of designing energy-aware inference workflows for intermittently powered systems. Future work will focus on calibrating real hardware operating thresholds and improving robustness to repeated power interruptions.

RELATED ABSTRACTS

Analysis of Large Language Model Watermarks
Presenter: Jonathan Rossignol
Faculty Sponsor: Lulu Kang
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C27]

Large Language Models (LLMs) are producing increasingly natural text which can be difficult to distinguish from human created text. One way to make LLM generated text identifiable is by applying a watermark in the generation process. Often this is done by introducing some form of pseudo-random noise into the logits produced by an LLM before applying the soft-max and sampling from the distribution of tokens. This makes the text identifiable using a detector function to check for this embedded pseudo-random noise which is unlikely to exist in human text.

While watermarks are an important step towards text provenance, they face certain limitations which makes them difficult to be effective in practice. Detector functions are prone to misidentifying both human and LLM generated text; watermarked text which has been even slightly modified by a human after generation may be difficult to identify; applying a watermark may degrade the text quality of the LLM. Here, I analyze and compare numerous watermarks such as the Green-list, Gumbel-max and SynthID to see how each counteracts the problems above. Using the Llama 3 LLM I generated text with and without watermarks using diverse prompts and varied token limits to be used as data for computing statistics such as perplexity and detection accuracy to evaluate each watermark's effectiveness.

RELATED ABSTRACTS

Toward LLM-Supported Automated Assessment of Critical Thinking Subskills
Presenter: Kushaan Naskar
Group Members: Payu Wittawatolarn, Brayden Liu
Faculty Sponsor: Andrew Lan
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 3, 1:15 PM - 2:00 PM: Room 163 [C28]

Critical thinking is a core competency in today’s education landscape, yet instructors still lack scalable ways to assess it and give students timely feedback. Our ongoing project explores whether we can automatically measure specific “subskills” that make up critical thinking in authentic student work. We focus on multiple data sources including student-written argumentative essays, and debates, where students synthesize sources, evaluate evidence, and use counterarguments.

We plan to develop a detailed coding rubric based on an established skills progression and use it for annotation across multiple critical thinking subskills. Building on this annotated dataset, we will investigate several approaches to automated scoring with large language models, including zero-shot prompting, few-shot prompting, and supervised fine-tuning, using both proprietary and open-source models.

In this poster, we will present our study design, rubric, and early qualitative insights from the annotation process, such as common challenges in distinguishing subskill levels and handling borderline cases. We will also outline our planned evaluation framework for comparing human and model-based scores and for examining where models succeed or struggle at the subskill level. Our long-term goal is to lay the groundwork for scalable, fine-grained assessment of critical thinking in student writing that can eventually support more targeted, pedagogically meaningful feedback in real classroom settings.

RELATED ABSTRACTS

Understanding Long-Tail Behavior in HNSW Recall
Presenter: Zacky Albandar
Group Members: Sevyn Gitau, Prateek Basavaraj
Faculty Sponsor: Marco Serafini
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 4, 2:15 PM - 3:00 PM: Room 165 [D16]

Large-scale AI systems rely on Approximate Nearest Neighbor (ANN) search methods to efficiently handle similarity queries. The Hierarchical Navigable Small World (HNSW) algorithm is currently the state-of-the-art approach for ANN search. While HNSW achieves highly efficient average-case performance, it exhibits suboptimal worst-case search complexity due to the long-tail distribution problem, in which a small percentage of queries account for a disproportionately large number of graph traversals or hops. This long-tail effect introduces significant delays in real-time applications. Despite this, existing research has yet to investigate HNSW’s intrinsic topological properties to determine the causes for the long-tail distribution phenomenon. To address this gap, our research provides a systematic analysis of the root causes of the tail distribution by evaluating the influence of key construction parameters on query performance. We conducted rigorous empirical investigations into the following three properties: data dimensionality, graph degree, and efConstruction, utilizing both synthetic data and standard high-dimensional benchmarks. Through our experiments, we hypothesize that by identifying the best parameters for graph degree, dimensionality, and efConstruction, we can shorten the long tail distribution, reducing the number of worst-case searches which take significantly more traversals to reach the top k nearest neighbors. Based on our experimental investigations, we expect to show that lower-dimensional embeddings and larger efConstruction settings will be the most effective in reducing the "curse of dimensionality”. Our findings will provide guidelines for optimal HNSW parameters, thereby facilitating the development of more cost-effective and efficient HNSW graphs for vector search applications. 

RELATED ABSTRACTS

Fortifying Small Language Models Against Query Injection Attacks
Presenter: Riddhimaan Senapati
Faculty Sponsor: James Allan
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 5, 3:15 PM - 4:00 PM: Room 163 [C20]

Relevance scores, a numerical representation of how useful a document is to a query, are used to calculate important information retrieval metrics such as precision, Normalized Discounted Cumulative gain etc. This process can be automated by Large Language Models (LLMs), but recent research has shown that LLMs are vulnerable to query injection where words from the query are injected into an irrelevant document leading to the LLM classifying the document as relevant. This thesis will focus on how to guard against these types of query injections for small language models (models with less than 10 billion parameters) by looking at prompt engineering techniques where we refine the prompt sent to the LLM, by running experiments with various type of attacks and mitigation strategies in order to find which techniques can help to make models perform better against query injection attacks. Results indicate that mitigation effectiveness varies significantly across different LLMs For gemma3:1b, few-shot prompting consistently performs best against all query injection attacks. For qwen3:0.6b, user prompt hardening generally achieves the best performance on all documents, though system prompt hardening performs better on relevant documents. These findings demonstrate that defending against query injection requires model-specific and sometimes even document-specific mitigation strategies.

RELATED ABSTRACTS

Mechanistic Interpretability Probing Information Retrieval Ranking LLMs
Presenter: Adriana Caraeni
Group Members: Romaisa Fatima
Faculty Sponsor: James Allan
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 5, 3:15 PM - 4:00 PM: Room 163 [C21]

This project investigates the internal representations of information retrieval (IR) features within the RankLLaMA 7B language model through layer-wise probing analysis. We extract neuron activations from all 32 layers of RankLLaMA when processing query-document pairs from the MS MARCO dataset and train Ridge regression models to predict around two dozen distinct IR features from these activations. Our feature set encompasses an extensions from traditional retrieval metrics (BM25, TF-IDF cosine similarity, KL and JS divergence), with term frequency features (min TF, normalized min TF, stream length), position-based features (proximity score, position bias, order preservation, term clustering), and advanced features (co-occurrence score, phrase matching, rare term score, query type score, semantic coverage, title boost, document length normalization, and TF saturation). By computing R^2 scores across layers, we identify where specific IR features are most strongly represented in the model's internal activations. Our results provide insights into how neural ranking models encode and process retrieval-relevant information across their layers, contributing to the mechanistic interpretability of language models for information retrieval tasks.

RELATED ABSTRACTS

Adaptive Re-Ranking
Presenter: Ata Cinar Genc
Group Members: Emir Kaan Korukluoglu
Faculty Sponsor: James Allan
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 5, 3:15 PM - 4:00 PM: Room 163 [C22]

Modern Information Retrieval (IR) systems typically use a "retrieve-then-rerank'' pipeline, where a computationally expensive, pre-determined cross-encoder re-ranks the top results from a fast initial retriever. While effective, this approach often applies heavy re-ranking models regardless of query complexity, resulting in high latency and wasted computational resources on simple queries. On our test set, only 40% of queries benefited from an accuracy gain with a re-ranker, and only 11% benefited from having a heavier re-ranker compared to a lighter one. We propose Adaptive Re-Ranking, a framework that dynamically routes queries to the most cost-effective strategy—ranging from sparse retrieval (BM25) and dense re-ranking (MiniLM-L6-v2) to heavy neural re-ranking (BGE-v2-m3)—based on query complexity. To train our routing classifier, we introduce a novel utility function to label queries based on their retrieval effectiveness (by combining nDCG@10 and MRR@10), penalized by latency. We curate a training dataset of over 300,000 labeled queries from diverse BEIR benchmarks. Compared to BGE, our method achieves around 1.15-53x lower median latency and around 1.11-3.52x lower mean latency across all datasets we have tested while delivering -19% to +5% effectiveness. Oracle results show that adaptive re-ranking can perform better than any fixed strategy. Our findings show that routing queries using our novel utility function offers a scalable solution to reduce computational costs and latency across a variety of IR systems.

RELATED ABSTRACTS

Bridging the Sim-to-Real Gap: Robust Vision-Based Navigation for Robotic Guide Dogs
Presenter: Shiven Patel
Faculty Sponsor: Donghyun Kim
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 5, 3:15 PM - 4:00 PM: Room 165 [D4]

Over 1.1 billion people worldwide live with some form of vision loss, yet fewer than 20,000 guide dogs are in service globally, leaving approximately 98% of eligible individuals without access to this gold standard of mobility assistance. Robotic guide dogs built on quadruped platforms offer a scalable alternative, but deploying reliable vision-based navigation in diverse, real-world urban environments remains a fundamental challenge. The primary bottleneck is the sim-to-real gap, the fidelity mismatch between simulated training environments and the physical world, which causes navigation policies to fail when transferred to real hardware.

This research proposes a real-to-sim pipeline that leverages 3D Gaussian Splatting and neural scene reconstruction to create photorealistic digital replicas of real urban environments from standard RGB camera footage. These reconstructions are integrated into NVIDIA Isaac Sim, where domain randomization with synthetic assets, pedestrians, vehicles, and street furniture, generates diverse training scenarios from a single captured scene. The approach combines the visual authenticity of real-world data with the scalability, safety, and controllability of simulation. Data collection will be conducted across the University of Massachusetts Amherst campus, capturing varied lighting, weather, and pedestrian conditions representative of typical guide dog deployment. 

The research contributes a complete pipeline from video capture through Gaussian Splatting reconstruction to physics-enabled simulation, a semantic domain randomization framework that tests cooperative sidewalk navigation skills, and a benchmarking environment for systematic evaluation of vision-based navigation policies. This work aims to demonstrate that encoding real-world structure into simulation can meaningfully reduce the sim-to-real gap for assistive robotic navigation.

RELATED ABSTRACTS

Building an Robust Accessibility-Object Dataset for Guide Dog Robots
Presenter: Duretti Diribaa Hordofaa
Group Members: Anshu Anjna, Dylan M. Gage
Faculty Sponsor: Donghyun Kim
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 5, 3:15 PM - 4:00 PM: Room 165 [D5]

While animal guide dogs provide essential mobility assistance to blind and low-vision (BLV) individuals, they are resource-intensive to train, expensive to maintain, and limited in availability. Furthermore, existing computer vision datasets lack the specific "accessibility objects” - such as door buttons, elevators, pedestrian signals, and crosswalks - necessary for training robust perception models on edge-compute platforms. Current datasets lack the quantity and quality of objects that are beneficial to navigation in the context of assisting BLV individuals. 
 
To bridge this gap, we introduce a novel, specialized dataset for accessibility-aware robotic navigation. We collected diverse imagery of accessibility objects across varying times of day, environmental conditions, and viewpoints. To enhance model robustness, we augment this real-world data using vision-language models (VLMs) to simulate varied environmental effects. We utilize Roboflow to generate precise segmentation masks and fine-tune lightweight segmentation architectures specifically for deployment on the Unitree Go2 quadruped robot. 
 
Our approach directly addresses the data scarcity problem in assistive robotics. By enabling real-time, on-device segmentation of critical navigation cues, we demonstrate a scalable pathway for autonomous guide robots. These results will support the development of guide dog robots to be more reliable and efficient in assisting BLV individuals navigate any environment. We expect this work to not only provide a foundational dataset for the community but also to validate real-world deployment. 
Task-Specific Robot Locomotion Policies Informed by Human Biomechanics
Presenter: Shlok Dharmendra Raval
Faculty Sponsor: Donghyun Kim
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 5, 3:15 PM - 4:00 PM: Room 165 [D6]

Reinforcement learning has enabled impressive locomotion behaviors in legged robots, yet most controllers remain agnostic to human biomechanics that could improve task-specific performance, such as energy efficiency and robustness. This project investigates how embedding biomechanical constraints from human motion into policy design can enhance robot performance on targeted motor tasks. I curate a library of canonical lower-limb motions, including level walking, stair ascent, and sit-to-stand from existing human biomechanics datasets and in-lab motion capture experiments. For each motion, I extract task-relevant constraints such as joint range envelopes, center of mass trajectories, and ground reaction force patterns and translate them into policy structure, state and action shaping, and reward terms for a simulated bipedal robot. Initial policies are obtained using a motion retargeting and imitation learning pipeline that maps recorded human trajectories onto the robot, after which I progressively introduce biomechanics-inspired constraints into the policy architecture, observation and action spaces, and reward design to study their impact on performance. I then train separate reinforcement learning policies for each motion using these constrained formulations and compare them against baseline policies trained without biomechanical priors. Performance is evaluated in physics simulation using metrics such as energy cost of transport, tracking accuracy, and disturbance recovery. Expected results are that biomechanically informed policies will achieve comparable or improved task success rates while reducing control effort and producing more human-like kinematics, providing a reusable pipeline for integrating human movement science into robotic control.
Driver Behavior Analysis Using CCTV Footage
Presenter: Drisilla Twumasi
Faculty Sponsor: Jay Taneja
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 6, 4:15 PM - 5:00 PM: Campus Center Auditorium [A45]

Each year, road traffic accidents continue to increase, with a majority attributed to unsafe driver behavior. Actions such as texting, speeding, or other forms of distraction place lives at risk whenever they occur. Although previous studies have examined the relationship between driver behavior and accident occurrence, limited research has focused specifically on developing cities such as Kigali, Rwanda. Historically, Kigali relied on diesel-powered buses that lacked advanced monitoring capabilities. The recent transition to electric buses now enables access to driver cabin CCTV footage, dashboard camera recordings, and telemetry data, allowing for improved monitoring of driver behavior. By analyzing these data sources, this study aims to identify associations between driver behavior and accident occurrences.


Driver footage provided by an electric bus company was analyzed to identify common behavioral distractions. The State Farm Distracted Driver Detection Kaggle dataset was used as a reference to guide activity labeling by identifying similarities between the recorded footage and known distracted driving behaviors. To support behavioral detection in video data, the deep learning-based Vision-Language Model AnomalyCLIP was applied to classify driver activities. Using global context optimization during training, AnomalyCLIP analyzes extracted video frames to categorize behaviors such as texting, drinking, eating, and self-grooming as abnormal behaviors.


Early findings indicate that driver behaviors can be classified into categories such as safe driving, texting, talking on the phone, drinking, eating, turning, self-grooming, improper seatbelt use, handling personal items, and interacting with passengers. This research seeks to develop accurate behavioral classifiers to support interventions aimed at reducing road accidents in Kigali.
Interpretable Semantic Axes for ‘Break' in Contextualized Embeddings
Presenter: Sujin Lee
Faculty Sponsor: Katrin Erk
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 6, 4:15 PM - 5:00 PM: Campus Center Auditorium [A64]

Large language models (LLMs) represent words as vectors in high-dimensional spaces called embeddings. Words that appear in similar contexts end up close together, forming geometric patterns that often mirror aspects of meaning. These patterns are remarkably systematic and thus the semantic terrain inside LLMs have long been investigated by researchers. Notably in the past Petersen and Potts (2023) have used the highly polysemous word break, with meanings ranging from literal physical breakage (“break a glass”) to emotional collapse (“break down”), and shown successful sense clustering in LLMs. However, I argue that clustering does not serve as sufficient evidence to answer what underlying semantic features drives the sense distinctions in LLMs and whether they align with meanings that humans recognize. My project aims to go beyond the prior findings of Petersen & Potts by implementing interpretable semantic axes, explicit directions in embedding space that correspond to psycholinguistic features. Using RoBERTa-large, I construct these axes from human feature-norm data (Binder et al., 2016) and project contextual embeddings of break onto them. I then analyze how different uses of the word separate along these dimensions. By implementing these axes to investigate whether an LLM “represents meaning” with cognitively real distinctions like causative vs. inchoative, my research connects computational patterns in LLMs with long-standing theoretical work in lexical semantics and cognitive science. 

Petersen, E., & Potts, C (2023), Lexical Semantics with Large Language Models: A Case Study of English “break”. Findings of EACL

Binder, J. R., et al. (2016), Toward a brain-based componential semantic representation. Cognitive Neuropsychology. Cogn Neuropsychol.

RELATED ABSTRACTS

Impact of Morphologically-Aware Tokenization on Language Model Performance Across Morphological Types
Presenter: Nathan Wolf
Faculty Sponsor: Katrin Erk
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 6, 4:15 PM - 5:00 PM: Campus Center Auditorium [A65]

Tokenization, the segmentation of text into a discrete sequence of “tokens”, is an essential yet often-overlooked step in all NLP tasks. Conventional tokenization schemes such as Byte-Pair Encoding (BPE) segment words without regard to their morphology. This lack of morphological information in tokens may hinder neural language models from learning an effective representation of a token’s meaning. Despite this, previous results do not conclusively show that incorporating morphological information into tokenization benefits language model performance. These mixed results could in part be due to not considering the morphological type of the languages on which the studies were conducted. To examine this, we compare how morphologically-aware tokenization impacts language model performance in fusional languages such as English, with relatively few morphemes per word, and agglutinative languages such as Finnish, with many morphemes per word. We hypothesize that conventional tokenization removes more morphological information in agglutinative languages than fusional ones, and thus morphologically-aware tokenization will help language model performance in agglutinative languages more than fusional languages. To investigate this hypothesis, we pretrain transformer language models on a multilingual web crawl dataset, which covers both fusional and agglutinative languages: English, Czech, Finnish, and Turkish. In each language, we pretrain one model with a conventional tokenizer, and one with a morphologically-aware tokenizer. We then evaluate all models on text generation, as well as a linguistically-motivated task. 
The Generative Generation? A Qualitative Investigation into Undergraduate AI Tool Use and Perspectives
Presenter: Lance Cheng
Faculty Sponsor: Ella Tuson
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 6, 4:15 PM - 5:00 PM: Room 163 [C26]

Over a short few years, large language models like ChatGPT have become ubiquitous academic partners-in-crime, annoyances, saving graces, fears, and everything in between. However, what everyday people think of these programs is frequently obfuscated behind extreme opinions from both sides. Additionally, these opinions often come with social pressure. In real life, professors may introduce no AI blanket terms for their classes that students cannot challenge or express their feelings on due to the power dynamic; students also may feel awkward opining around their classmates due to general peer pressure. Online, tech startup culture is extremely prevalent on career-focused websites such as LinkedIn, and has further spread to “mainstream,” general-purpose social media sites, especially Instagram and Twitter; influencers in this sphere depict themselves using AI for all kinds of everyday tasks, and are often themselves developing AI tools. Due to these social dynamics, which often result in mutual mistrust, there is a severe need for clarity on young people's true thoughts and motivations to use AI. Through a series of interviews, this study seeks to characterize UMass CICS undergraduates' use of LLMS and other generative AI tools, in particular their motivations for using or not using them, with an eye towards the role of complex academic social dynamics. The researchers hope this study will provide recommendations to CICS staff on how to move forward crafting an educational experience with AI awareness and a student-first lens.

Why Technically Advanced Products Fail: A Study of User Interface, Social Acceptability, and Product Adoption Across Emerging Technologies
Presenter: JAHNAVI SHARMA
Faculty Sponsor: Ella Tuson
School: UMass Amherst
Research Area: Computer Science
Location: Poster Session 6, 4:15 PM - 5:00 PM: Room 163 [C27]

Emerging technologies often arrive surrounded by bold claims about transformation, efficiency, and the future of everyday life. Still, many never move beyond early curiosity. The gap between what a product promises and how people actually live with it points to a deeper design challenge. How does a technology move from demonstration to habit? This study investigates how early design choices shape that trajectory.

Focusing on wearable augmented reality systems such as Google Glass, spatial computing platforms like Apple Vision Pro, social smart glasses including Ray-Ban Meta Smart Glasses, and immersive metaverse environments such as Meta Horizon Worlds, this research explores how interface design, embodied interaction, and public presence influence long-term engagement. Each case reflects a different attempt to bring digital systems directly into shared social space.

The methodology combines close interface analysis with a synthesis of early adopter reviews, developer materials, and media coverage. Cognitive complexity, onboarding flow, and task structure were assessed in relation to patterns of sustained use and disengagement. Early signals were explored through online discussions and public conversations to understand general reactions such as excitement, hesitation, or curiosity. These observations were considered alongside broader patterns like public reception, engagement levels, and shifts in how products were positioned over time.

Across cases, similar pressures surface: unclear everyday purpose, social unease during visible use, and initial excitement that fades without lasting relevance. Adoption unfolds through lived experience, shaped by how naturally a system fits into daily routines and shared spaces. This project develops a user-centered framework for identifying early signals of traction, offering practical guidance for designers, founders, and investors navigating high-stakes innovation. At a time when emerging technologies rapidly reshape public space, work, and social interaction, understanding how design choices translate into cultural acceptance is essential to building systems that endure.