Physical exercise is vital for maintaining health and well-being, and it is equally important for individuals who are blind, as it supports not only physical fitness but also contributes significantly to functional mobility, independence, and quality of life. However, for exercise to be safe and effective, it must be performed with proper technique, which can be especially challenging for blind people due to lack of visual feedback. Blind individuals may struggle to align their posture, control their movements, or follow exercise routines accurately, increasing the risk of injury or reduced effectiveness.

To overcome these challenges, blind people need specialized support, preferably in the form of AI-enabled assistive technology that can provide clear real-time audio guidance to ensure safe and effective physical activity. In this project, the student will develop such assistive technology in the form of a mobile app which will leverage the smartphone camera to observe the physical activity of a blind person and then use customized large language models to analyze if the exercise is being done properly. If needed, the app will provide real-time audio instructions, e.g., for posture correction, to assist the blind person.

Student Learning Outcomes:

The student will learn how to build and analyze multimodal datasets, fine-tune open-source large language models (LLMs), design LLM prompts, develop AI-powered smartphone apps, design non-visual usable interfaces, and lastly, analyze and interpret performance data from pilot evaluations.

Skills Needed:

Python, Android, Java, Gemini.

This project explores how individuals collaborate with AI systems to evaluate the credibility and clarity of health-related content in digital environments, with a specific focus on influencer-generated posts about Attention Deficit Hyperactivity Disorder (ADHD). The research centers on understanding how humans team with AI assistants that provide real-time feedback, such as summaries, source cues, or warnings, while reviewing content on platforms like TikTok, Instagram, or YouTube. By investigating how this AI support affects user trust, attention, and judgment, the study contributes to emerging research on human-AI teaming in health information systems.

The student will design and conduct a controlled eye-tracking experiment in which participants interact with a web-based interface that presents ADHD-related content collected from publicly available social media posts. This interface will simulate an AI teammate that provides augmented information, such as explanatory text, content credibility indicators, or risk warnings generated using natural language processing tools. As participants engage with the content, their visual attention will be tracked using eye tracking. The study will generate data from multiple sources, including participant gaze patterns, click behavior, and self-reported feedback on trust and content understanding. These data will be analyzed to identify patterns in how AI explanations influence attention distribution and decision-making behavior.

Student Learning Outcomes:

The student will gain experience designing and running user studies, working with AI-generated content, and analyzing human attention data using Python and visualization tools. The student will develop a functional prototype, collect and interpret results from a pilot user study, and prepare a written research report that can serve as a foundation for future academic publications.

Skills Needed:

Python, PyPlot.

This project investigates how different styles of image explanations generated by Vision-Language Models (VLMs) impact human understanding, trust, and perceived helpfulness. The focus will be on comparing explanation types, such as concise vs. detailed, neutral vs. emotionally expressive, and technical vs. simplified, and their effect on users’ interpretation. We also study whether VLMs generate effective image descriptions for complex web content such as infographics and diagrams.

Students will begin by curating a broad dataset of web images, including news photos, medical images, product visuals, infographics, and diagrams. Using state-of-the-art multimodal LLMs (e.g., GPT-4V, Gemini, Claude 3), students will prompt the models to generate image captions in varying styles. Sentiment analysis will be applied to characterize the tone and emotional framing of the captions. A user study will then be designed and conducted to evaluate how different explanation styles influence users’ perceived trust, comprehension, and helpfulness. Finally, the study results will be analyzed and compared to the sentiment characteristics of the AI-generated captions, uncovering relationships between explanation tone and user perception.

Student Learning Outcomes:

Students will understand the fundamentals of VLMs, build a dataset, write scripts for a prompt-based VLM experiment, apply sentiment analysis, design a user study, and analyze the experiment results.

Skills Needed:

Python, vision LLM prompting, user study, data analysis.

With the rapid advancement of LLMs, AI has been used for generating virtually any type of text, including science news. However, how well this generated news helps human readers to gain scientific knowledge and how well AI-generated news compares with human-written news are still open questions. In the REU projects, we will investigate three scientific questions:

1. What is a good evaluation metric that accounts for the knowledge gained by human readers?
2. How effective are AI-generated compared with human-written science news in terms of helping human readers to gain knowledge?
3. How to improve the quality of AI-generated news to increase the knowledge gained by human readers?

We will perform our experiments based on a website we have built that allows humans to read science news and answer questions. We will perform human studies on undergraduate and graduate students. The human-written science news is adopted from MSVEC2.0, our carefully curated science news dataset.

Student Learning Outcomes:

The student will learn text evaluation metrics, how to design experiments to verify the alignment of automatic metrics vs. human evaluation metrics, various prompt engineering techniques to generate science news, and how to perform human studies, analyze, and visualize data.

Skills Needed:

Python, web programming, database management.

The objective of this project is to develop and validate an automated writing evaluation (AWE) system capable of classifying argumentative elements in persuasive essays written by students in grades 6–12 across the United States as “effective,” “adequate,” or “ineffective.” The study will leverage the publicly available PERSUADE 2.0 corpus to conduct secondary data analysis. The analysis will exclusively target essays written by students who were receiving Individualized Education Programs (IEPs) at the time of data collection, with the AWE system designed to account for their unique writing characteristics and learning needs. The models derived from this work will lay the foundation for an LLM-driven learning tool designed to support struggling writers by offering automated, targeted feedback to enhance their persuasive writing proficiency.

Student Learning Outcomes:

The student will gain hands-on experience in designing and fine-tuning LLMs for educational applications by working with written response data using computational methods. During the nine-week program, the student will analyze and interpret findings from the secondary data analysis and produce a written research report that may serve as the basis for future academic publications.

Skills Needed:

Python.

The goal of this project is to:

1. identify evaluation categories and features of standardized patients interacting in Objective Structured Clinical Examinations (OSCEs) scenarios with an emphasis on quality-of-service and accurate portrayal;
2. assess the effectiveness and efficiency of the ML and neural network models; and
3. compare AI assessments with human-generated evaluations for validation.

Data will consist of sample sessions of medical students and standardized patients.

Student Learning Outcomes:

The student will develop a deeper understanding of ML algorithms, algorithm performance metrics, data labeling and annotation, and topic segmentation.

Skills Needed:

Python, fundamentals of ML and neural networks.