Joshua West

Joshua West

Biography

Joshua West is a rising sophomore Honors Electrical Engineering student at North
Carolina Agricultural and Technical State University and a research intern at the Massachusetts
Institute of Technology. His research focuses on artificial intelligence, machine learning, and the
development of consistent behaviors in large language models. During his first year at NC A&T,
he conducted undergraduate research using Python and environmental sensor data to study urban
microclimates in Greensboro. He is passionate about using technology to solve real-world
problems and create opportunities for underserved communities. Outside of research, Joshua is the
co-founder of Champions of Change. This nonprofit organization has impacted more than 10,000
people across seven states and 10 cities through literacy initiatives, community outreach, and youth
leadership programs. Joshua aspires to become an entrepreneur and engineer who develops
innovative technologies that create lasting change in communities around the world.

Probing Language Model Accuracy Through Behavioral Evaluation and
Representation Analysis
Joshua West1, Khai Pham2, Laura Ruis3, and Jacob Andreas3
1Department of Electrical and Computer Engineering, North Carolina Agricultural and Technical
State University
2Department of Computer Science, Rice University
3Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology

Large language models are trained on massive datasets, and that scale is exactly what makes
them useful, but it also means they may piece together relationships that were never explicitly
stated. Prior work has tested whether models can recover a single hidden fact, like an unknown
city, from scattered clues. Real-world knowledge, though, rarely comes as an isolated fact. It comes
as a network of relationships. This project asks: given only partial, indirect evidence, what can a
model infer about a graph of connections it was never shown, and how does that depend on graph
complexity? We fine-tuned Qwen3-0.6B on synthetic relational graphs across four complexity
configurations, evaluating its ability to predict entities and relationships on held-out data. Accuracy
followed a non-monotonic, U-shaped pattern, ranging from roughly 46% at one mid-complexity
configuration to nearly 74% at the highest. That mid-complexity configuration underperformed
even simpler ones on relationship classification, not just entity naming, and the dip didn’t track with
training loss. It remains unresolved. This unexplained dip is the real finding. Can a model’s internal
representations predict this kind of unreliable behavior before it surfaces? The implications extend
beyond this dataset. The same mechanism could affect how models reason about sensitive relational
knowledge, from medical records to dangerous technical processes