UVA School of Data Science: Capital One Fellows 2025-2027

Capital One has a long-standing partnership with the University of Virginia’s (UVA) School of Data Science, which has continued to result in some exciting advancements, including the opening of the Capital One Hub in 2024.

Building on this collaborative spirit, we’re proud to announce the recipients of the 2025-2027 Capital One Fellowship Awards. These awards provide School of Data Science faculty with support for two doctoral candidates, named Capital One Fellows, for a period of two years. The awards will enable faculty to conduct advanced research in the critical fields of artificial intelligence (AI) and machine learning (ML). We look forward to the innovative contributions these fellowships will yield.

Award Recipients 2025-2027

Sheng Li is an AI researcher, and his long-term goal is to develop intelligent systems in open and dynamic environments. Li joined the UVA School of Data Science in 2022.

Prior to UVA, Li was an assistant professor of computer science at the University of Georgia from 2018 to 2022 and a data scientist at Adobe Research from 2017 to 2018. He currently directs UVA’s Reasoning and Knowledge Discovery Lab. Li’s research interests include trustworthy ML (e.g., robustness, fairness, causality, transferability), generative AI (e.g., large language models [LLMs], diffusion models), computer vision and causal inference.

Project: Integrative Decoding for Reliable LLM Reasoning in Financial Services

As LLMs are increasingly deployed in high-stakes domains such as finance, health care and education, ensuring that their outputs are trustworthy and factual is no longer optional. For example, LLMs often suffer from hallucinations, generating plausible-sounding yet incorrect statements, which can lead to costly or dangerous decisions. Mengxuan Hu’s prior research primarily focused on improving the safety alignment of LLMs, particularly in high-stakes scenarios. Building on this foundation, she is excited to extend her work into the financial domain through this project. In particular, Hu is drawn to the ID framework, which provides a principled way to promote self-consistency and factuality in LLM outputs without requiring model retraining. She is especially interested in developing reinforcement learning-based decoding policies tailored to financial tabular reasoning tasks such as fraud detection. She looks forward to exploring how decoding-time interventions, guided by domain-specific reward functions, can robustly improve both factual accuracy and decision quality in real-world financial applications.

This two-year project aims to enhance the accuracy and efficiency of LLMs in financial services, particularly for tasks involving tabular data like fraud detection. LLMs can “hallucinate” false information, a critical issue in financial domains. Integrative decoding (ID) is a promising solution that improves factual accuracy by incorporating implicit self-consistency during text generation, sampling multiple outputs and selecting tokens based on aggregated predictions. ID maintains efficiency by processing samples in batches, adding minimal latency.

Chirag Agarwal is an assistant professor of data science and leads the Aikyam Lab, which focuses on developing trustworthy ML frameworks that go beyond training models for specific downstream tasks and satisfy trustworthy properties, such as explainability, fairness and robustness.

Before UVA, Agarwal was a postdoctoral research fellow at Harvard and earned his doctorate in electrical and computer engineering from the University of Illinois at Chicago. His doctoral thesis focused on the robustness and explainability of deep neural networks. His research covers explainability, fairness, robustness, privacy and transferability estimation in large-scale models. He developed a pioneering, large-scale study for systematic, reproducible and efficient evaluations of post hoc explanation methods for structured and unstructured data, aimed at understanding algorithmic decision-making in diverse tasks.

Project: Toward multimodal graph-language model reasoning

Graph data is ubiquitous in real-life scenarios, playing a central role in numerous scientific disciplines. In biology, for instance, complex molecules and the relationships between biological entities can be naturally represented as graphs. Modeling graph data is essential for understanding molecular interactions or predicting biological behaviors. Ding Zhang is interested in modeling graph data using graph representation learning techniques, e.g., applying graph neural network models.

Additionally, as AI systems become increasingly integrated into critical decision-making processes, such as health care, finance and public policy, it is essential that these systems are not only accurate but also interpretable and reliable. Building models that can clearly explain their predictions is crucial for fostering user trust and ensuring ethical deployment. This area is key to bridging the gap between technical innovation and the real-world, responsible adoption of AI.

This project proposes a new benchmark with modal-complementarity tasks, forcing GLMs to integrate graph and text data for complex reasoning (e.g., fraud detection). Graph-language models (GLMs) can struggle with outdated evaluation and overreliance on single data modalities, especially in fields like finance. Additionally, the Modal Fusion Adapter framework dynamically balances the influence of graph and text embeddings, preventing overreliance. This approach aims to unlock GLMs’ full potential in real-world applications.

Advancing AI research in financial services

Capital One’s ongoing partnership with UVA’s School of Data Science, highlighted by these awards, underscores the vital role of collaboration between industry and academia. By supporting cutting-edge research in AI and ML, these fellowships not only foster innovation and advance the field but also equip the next generation of data scientists with the skills and knowledge to address complex challenges in real-world applications.