Computer Science

Session I

Linwood Bailey, "Evaluating the Generalizability of Scene-Flow Specifications: Real-World Implementation for Traffic Law Monitoring on ROSbot XL"

Ensuring that autonomous vehicles abide by regional driving laws requires bridging raw perception with formal temporal logic. Scene-Flow Language addresses this challenge by synthesizing scene graphs, finite automata, and temporal logic (different ways to write and evaluate logical expressions) to monitor system violations. To validate the practical generalizability of Scene-Flow beyond simulation, I implemented a real-time compliance monitor on a ROSbot XL autonomous mobile robot. I encoded a rule targeting improper overtaking maneuvers found in Virginia Traffic Code § 46.2-838. To feed clear data into the monitoring framework, I developed a perception pipeline coupling LiDAR point-cloud clustering with a pre-trained OpenCV bounding box detection model. The successful end-to-end deployment confirms that Scene-Flow can be seamlessly deployed on physical ROS 2 hardware architectures, validating its utility as a generalized framework for autonomous traffic safety enforcement. 

Student Major(s)/Minor: Computer Science Major

Advisor: Dr. Trey Woodlief

Dylan Han, "Tribe Street View: Using 3D Gaussian Splatting to Create a Walkable Virtual Archive of Campus"

3D Gaussian Splatting, a computer graphics technique that creates 3D scenes from images, has recently come into prominence for its quick training time and rendering speed when compared to other AI rendering techniques, which are computationally more expensive to train and evaluate. Part of its promise is its ability to create continuous, navigable landscapes, rather than the discontinuous scenes from tools like Street View which feature large jumps between viewable regions. This discontinuity can be especially unsatisfying in contexts such as virtual tours, where jumps limit the viewer from experiencing a richer view of the venue. This project seeks to use 3D Gaussian Splatting to create a virtual archive of campus for the preservation of its appearance at this point in time and for viewing by the general public. This research has the potential to democratize campus tours, providing prospective students from disadvantaged backgrounds the opportunity to explore campus virtually. 

Student Major(s): Computer Science and Finance Major

Advisor: Dr. Pieter Peers

Arden Michel, "Towards a safer future: Creating a scalable quantum network using BB84 and quantum routers"

This study aims to replicate and localize the 2025 study Broadband internet access as a social determinant in the early COVID-19 pandemic in US counties by Spencer Allen to the Hampton Roads area for reproducibility as well as apply the study with more robust broadband internet access data and standards. Data was drawn from publicly available datasets and analyzed using OLS regression. Results show a negative correlation between COVID-19 outcomes and mask usage with broadband internet access. Thus, the index study rings true under scrutiny, and as we move towards a more digitized world, broadband internet access can be viewed as a determinant for public health. 

Student Major(s): Computer Science Major

Advisor: Dr. Qun Li

Session II

Liang Huang, "LLM-Generated Test Oracles for Evaluating Security in Quantized Code Models"

Full-precision code models are costly to serve at scale, so practitioners quantize them. Whether that trade-off costs security had gone untested. Static analyzers flag patterns; testing this needs a test oracle that observes whether code resists exploitation. This project built and validated one: SecurityEval-Exec, an execution-based version of SecurityEval whose test suites were generated by Claude Opus 4.6. Checked against CWEval’s independent human oracles across 3,089 programs, the generated oracles agreed with human verdicts 95.5% of the time, accepted secure references 80.8% of the time, and detected 90.0% of vulnerabilities. Using both benchmarks, we evaluated Qwen2.5-Coder-Instruct (7B/14B/32B; full precision, 4-bit AWQ, and GPTQ) under three system prompts across 159 tasks, 51 CWE types, and five languages. Quantization produced no systematic security penalty; security-aware prompting helped unevenly. The lesson: a test that runs is not the same claim as one that is right. Only independent validation tells them apart. 

Student Majors: Computer Science and Mathematics Major

Advisor: Dr. Antonio Mastropaolo

Xingyue Liu, "From Scrolling to Scholarship: Designing a Dual Computational Model for Academic Self-Discovery"

Undergraduate academic indecision causes significant major switching, particularly across quantitative STEM disciplines. Traditional academic counseling faces severe capacity limits and short-term efficacy. To provide scalable, continuous guidance, this project introduces a Dual Computational Model that repurposes deep neural network recommendation architectures for ethical academic self-discovery. Stage 1 utilizes a Two-Tower neural network and topic diversification algorithms to map unstructured course descriptions and broad student interests into a shared embedding space while actively dismantling departmental filter bubbles. Stage 2 executes high-precision ranking by capturing real-time user telemetry—including cursor dwell times and scrolling speeds—to resolve query ambiguity. Evaluated via a web-facing sandbox using privacy-preserving synthetic profiles, the framework effectively down-ranks redundant departmental listings while aligning with specialized academic niches. This zero-marginal-cost platform empowers students to navigate from digital scrolling to authentic scholarship. 

Student Majors: Computer Science and Mathematics Major

Advisor: Dr. Yuchen Wang

Kevin Wang, "Constraining the Deputy: Resource-Aware information Flow Control at the MCP Boundary"

Model Context Protocol (MCP) agents may use one file-reading tool for an untrusted email and a confidential record. This project asks whether an MCP gateway can enforce confidentiality at the resource level without modifying agents, models, or tools. I developed MCP-IFC, an information-flow-control monitor that derives security labels from administrator-defined file locations, tracks access to protected data, and evaluates subsequent operations. I tested seven policies across 42 indirect prompt-injection scenarios, 415 organizational emails, and two language models. Unmonitored attack success ranged from 5.1 to 89.7 percent. Resource-aware enforcement contained every observed write-based leak. The strictest policy blocked 7.0 percent of legitimate tasks; avoiding those blocks required trusted task authorization before execution. Disclosures occurring outside the MCP boundary therefore require complementary agent-level security controls. The results show how resource-level labels distinguish operations that fixed tool labels collapse while defining the role of boundary enforcement in a layered defense. 

Student Major(s)/Minor: Computer Science Major

Advisor: Professor Yue Xiao

Video Presentation