Research
Learning, reasoning, and agency in humans and AI
Research Overview
I study learning, reasoning, and decision-making in humans and AI from a statistical perspective. My research develops statistical ideas for understanding modern AI systems and studies how those systems can be designed to support human judgment and agency.
AI as Probabilistic Computing
I view modern AI as a form of learned, probabilistic computing. Rather than specifying every behavior explicitly, we use data to construct complex systems whose representations, predictions, and decisions emerge through learning. Understanding such systems raises fundamental statistical questions about dependence, uncertainty, generalization, measurement, and decision-making.
My research asks these questions at two levels: what we can infer about AI systems themselves, and what we can infer about and support in the humans who interact with them.
My work spans both machine and human learning. I develop statistical and information-theoretic methods for understanding learned representations and generalization in AI systems, and I develop new approaches to measuring how people learn, reason, and make decisions through interactions with AI. Across both, I am interested in how uncertainty, information, and statistical evidence should inform decisions without unnecessarily replacing human judgment or agency.
I collaborate with researchers and organizations in areas where these questions matter in practice, including education, family history and genealogy, cybersecurity, and scientific discovery.
Research Areas
Understanding AI Systems
What do complex learned systems learn, and what properties persist?
Modern AI systems learn high-dimensional representations whose structure is poorly understood. Many classical inferential tools focus on relatively low-dimensional parameters or functionals. Modern AI instead asks us to reason about complex fitted objects such as learned representations, latent geometries, dependence structures, and generated distributions, motivating new approaches to statistical inference. I develop statistical and information-theoretic methods for characterizing what these systems learn and which properties are stable.
Current work includes Local CorEx, which uses total correlation to discover higher-order dependence structure in neural network representations, and studies of diffusion model generalization that ask what statistical regularities persist across independently trained generative models.
Human-AI Learning and Measurement
What can we infer about human knowledge and reasoning from rich interactions with AI?
Many traditional assessments measure complex human competence through responses to a relatively small set of fixed items. AI enables adaptive, interactive measurement in which the evidence can include not merely whether a person selected the right answer, but how they recognize a situation, gather information, reason through alternatives, respond to feedback, and make a decision. I study how to turn those rich behavioral traces into valid statistical evidence about a person's underlying knowledge and capabilities.
This is fundamentally a measurement problem. What is the estimand? What evidence identifies it? How reliable is the measurement? How much does the interaction policy affect the result? How do we separate ability from nuisance variation? Current projects explore AI-administered oral assessments, decision-based simulations, and experiments on how people learn and reason with AI.
Trustworthy Human-AI Systems
How should uncertain information be used when humans and AI make decisions together?
Prediction has never really been the endpoint of statistics—ultimately someone makes a decision. AI makes this boundary unusually visible because the computational system may itself retrieve information, withhold it, recommend, classify, or act. My work explores how uncertainty, information, risk, and competing objectives should shape these decisions, drawing especially on statistical decision theory and ideas about utility and information.
A further question motivates much of this work: which decisions should we optimize AI to make, and which should we optimize AI to help humans make? This is a natural modern extension of statistical decision theory that incorporates human agency, trust, and appropriate delegation. Current projects include RedactaRAG, which studies how policy constraints and information boundaries should interact with retrieval and generation in LLM systems.
Application Domains
I collaborate with researchers and organizations in areas where these questions matter in practice. These are application domains and collaborations, not separate research programs.
Education
AI-assisted learning, assessment, and simulation
Family History & Genealogy
Graph-based methods for assisted family tree construction
Cybersecurity
Anomaly detection, network security, policy-aware systems
Scientific Discovery
Molecular property prediction, bioinformatics
Publications
Dissertation
Graph-Based Machine Learning: Higher-Order Interactions, Guided Generation, and Knowledge-Graph Tools
PhD Dissertation, Utah State University, 2025
Advisor: Kevin Moon
This dissertation develops novel methodologies for learning on graph-structured data, with contributions to higher-order interaction modeling, guided graph generation, and practical knowledge-graph tools for real-world applications.
Working Papers & Manuscripts in Preparation
Hardy, T, Kerby, T
Working Paper (2026)
Introduces RedactaRAG, a policy-aware retrieval and redaction framework for secure LLM workflows, and studies convergence behavior between policy constraints, retrieval quality, and generation utility in real-world deployment settings.
Kerby, T, Moon, K.
TBD (2026)
Journalizes "Learning Local Higher-Order Interactions with Total Correlation" and adds additional Neural Network experiments.
Papers Under Review
Adam G. Rustad, Jun Kim, Jooyoung Jeon, Thomas Kerby, Jake Slater Rhodes
NeurIPS (submission) (2026)
Independently trained diffusion models often produce nearly identical images from the same initial noise under deterministic DDIM sampling; we study when agreement fails. Early steps contract trajectories into a shared low-dimensional region, while later steps apply model-specific refinement.
Peer-Reviewed Publications
Kerby, T, Fuller, B, Moon, K.
JOSS (2026)
Presents Nexarag, an open-source, containerized platform that combines citation graphs, semantic embeddings, and retrieval-augmented generation to help researchers interactively explore, synthesize, and identify gaps in rapidly growing scientific literature.
N. Brimhall, K. Bladen, T. Kerby, C. Legleiter, C. Swapp, H. Fluckiger, J. Bahr, M. Roberts, K. Hart, C. Stegman, B. Bean, K. Moon
Remote Sensing (2026)
Introduces a large, publicly available dataset of over 250,000 georeferenced river images from across the U.S., along with a computer vision pipeline for automated rapid detection, enabling scalable remote sensing applications in riverine science and resource management.
Kerby, T, White, T, Moon, K
IEEE 34th International Workshop on Machine Learning for Signal Processing (MLSP) (2024)
Novel method for capturing higher-order interactions using total correlation and demonstrates its effectiveness on synthetic and real-world data including the internal representations of neural networks.
Other Publications
Kerby, T, Moon, K
ArXiv (2024)
Introduces a framework for training-free guidance in discrete diffusion models and demonstrates its effectiveness in guiding molecular graph generation using DiGress.
Butler, T, Frandsen, A, Lightheart, R, Bargh, B, Taylor, J, Bollerman, T, Kerby, T, West, K, Voronov, G, Moon, K, Kind, T, Dorrestein, P, Allen, A, Colluru, V, Healey, D
ArXiv (2024)
Introduces MS2Mol, a generative transformer model for de novo molecular structure prediction from mass spectra, and demonstrates its superior performance on a new benchmark dataset of unknown natural products.
Collaboration
I welcome collaborations with researchers and organizations whose work intersects with the statistical and human-centered questions above. I am especially interested in partnerships that combine methodological innovation with real-world impact.
Interdisciplinary Research
Combining statistical and machine learning methods with domain expertise across education, scientific discovery, cybersecurity, and family history—particularly where questions of measurement, uncertainty, and decision-making are central.
Industry Partnerships
Real-world applications where trust, interpretability, and human agency are central concerns, and where statistical thinking can contribute to building reliable AI systems.
Research Resources
- • Software: PyTorch, NetworkX, LangChain
- • Computing: GPU clusters, distributed computing frameworks
- • Data Sources: Public datasets, collaborative research data, benchmark collections
- • Tools: Custom data science libraries (available on GitHub)