Dmytro Vitel

Postdoctoral Researcher, PALM Lab
Bellini College of Artificial Intelligence, Cybersecurity and Computing, University of South Florida

About

I am a postdoctoral researcher at the PALM Lab (Pioneering Advancements in Learning Methods) in the Bellini College of Artificial Intelligence, Cybersecurity and Computing at the University of South Florida in Tampa, working with Dr. Anshuman Chhabra on language model behaviour and interpretability. I completed my Ph.D. at USF in 2026 under Dr. Alessio Gaspar in the CEREAL Lab, where my dissertation studied how latent structure in semantic spaces can be extracted and exploited for efficient symbolic search and black-box optimization.

My research interests centre on the study of language models: how their behaviour arises from training data and internal representations, where they fail, and how they can be made more reliable. In training-data influence estimation, I have studied how the choice of layers and the aggregation of influence across layers affect the identification of beneficial and detrimental training samples in large language models. My current direction of study is the model collapse problem under recursive training, together with agentic AI and world modeling within large neural models.

My earlier research concerned structured search in learning systems: symbolic search, semantic genetic programming and interaction-aware optimization in evolutionary computation, where extracting the latent interaction structure of a search space improves convergence and mitigates premature solutions.

Research interests

Language modeling, model interpretability and training-data influence
Agentic systems and world modeling
Symbolic search, semantic genetic programming and evolutionary computation
Coevolution and latent-structure extraction from sparse interactions

Selected publications

First is Not Really Better Than Last: Evaluating Layer Choice and Aggregation Strategies in Language Model Data Influence Estimation Dmytro Vitel, Anshuman Chhabra International Conference on Learning Representations (ICLR), 2026 paper arXiv code data
Coordinate System Extraction as the Search Driver in Test-Based Genetic Programming Dmytro Vitel, Kok Cheng Tan, Alessio Gaspar, Paul Wiegand Genetic and Evolutionary Computation Conference (GECCO), ACM, 2025 paper code
Robust Search for the Underlying Objectives in Black-Box Games with Binary Outcomes Dmytro Vitel, Alessio Gaspar, Paul Wiegand Applications of Evolutionary Computation (EvoApplications, EvoStar), LNCS, 2025 paper code

Full list on Google Scholar and in the CV.

Education

Ph.D. in Computer Science, University of South Florida, 2026: Dissertation: Extracting and Exploiting Latent Structures in Semantic Spaces for Efficient Symbolic Search and Black-Box Optimization. Advisor: Alessio Gaspar.
M.S. in Computer Science, University of South Florida, 2019: Thesis: On Automation of Code Transformations for Concept Inventory Misconceptions. Advisor: Alessio Gaspar.
M.S. in Applied Physics, Taras Shevchenko National University of Kyiv, 2015: Thesis: Parallel Routing Algorithms on Graphs for Heterogeneous Architectures. Advisor: S. D. Pogorilyy.

Contact

dvitel [at] usf [dot] edu · Tampa, FL, USA