Research
Our group develops statistical and machine learning methods for understanding representation, reasoning, and decision-making in complex systems.
A common challenge across the human brain, modern AI, and financial markets is that decisions must be made from information that is high-dimensional, noisy, dynamic, and incomplete. We are interested not only in predicting what happens next, but also in understanding how information is represented, how uncertainty evolves, and how representations lead to reasoning and decisions.
Our research combines statistical methodology, machine learning, geometry, generative models, and sequential decision-making. We currently pursue these questions in three interconnected areas: human cognition and neuroscience, AI reasoning and agents, and financial decision-making.
1. Human Cognition and Brain Networks
Our longstanding research in neuroscience develops statistical and machine learning methods for understanding the organization, variability, and adaptation of the human brain. A major focus is the human connectome—the structural and functional connections linking different parts of the brain. Unlike conventional approaches that reduce the brain to a fixed collection of regions, we develop methods that preserve its continuous geometry and complex network structure.
Current directions include:
- Continuous and geometry-based connectome analysis — developing atlas-free statistical representations of structural and functional brain connectivity.
- Multimodal brain modeling — integrating structural connectivity, functional connectivity, imaging, genetics, and behavioral measurements.
- Generative models for neuroimaging — using latent-variable and diffusion models for harmonization, reconstruction, and prediction.
- Brain aging and disease — identifying structural and functional changes associated with aging, cognitive decline, pain, and neurological conditions.
- Cognitive learning and resilience — studying how the brain adapts to challenging cognitive tasks and how personalized interventions may strengthen cognitive function.
- Reproducible computational neuroscience — building statistical and computational tools that enable large-scale collaborative studies of brain connectivity.
This research is supported by the National Institutes of Health (e.g., MH118927, AG066970, R25DA058940), including projects on brain connectivity, aging, cognitive training, and substance-use research.
2. Reasoning, Representation, and AI Agents
Large language models increasingly operate not simply as predictors of text, but as reasoning and decision-making systems. We study how these models represent information internally, how reasoning trajectories develop during generation, and how reliably those trajectories translate into decisions. A central question is:
Given the same information, when and how does an AI system commit to a particular reasoning path or decision?
Current directions include:
- Internal representations and reasoning trajectories — studying how high-dimensional model representations evolve as reasoning and decisions emerge.
- Uncertainty and robustness — quantifying how reasoning paths and final decisions vary across repeated generations, perturbations, and alternative contexts.
- Sequential and agentic decision-making — studying LLM-based agents that repeatedly observe, reason, act, and adapt within an environment.
- Reasoning with tools — integrating language models with retrieval systems, simulations, optimization algorithms, statistical models, and other external tools.
- Learning from outcomes — developing training and adaptation strategies that optimize long-term behavior rather than individual predictions.
- Interpretability and reliability — identifying when internal representations become predictive of subsequent reasoning, actions, and failures.
Our broader goal is to develop statistical frameworks for understanding AI systems as stochastic, adaptive decision-making processes, rather than treating them solely as static prediction models.
3. Financial Intelligence and Decision-Making
Financial markets provide a natural environment for studying decision-making under extreme uncertainty. Prices, news, economic information, and the behavior of other market participants interact dynamically, creating a challenging setting for statistical learning and AI. Our work combines predictive modeling with decision and risk analysis.
Predictive modeling and information extraction
We develop models that integrate heterogeneous sources of financial information, including market prices and trading signals, financial and economic time series, news and unstructured text, and representations extracted by large language models. We are particularly interested in understanding when predictive signals are reliable and when uncertainty should prevent a model from acting.
Sequential decision-making and AI agents
We study AI agents that repeatedly observe financial information, form beliefs, and make decisions over time. Research questions include:
- Uncertainty-aware trading and decision rules
- Information aggregation across models and data sources
- Portfolio optimization and dynamic risk management
- Evaluation of decision strategies under distributional change
- Robustness and generalization of financial agents
- Training agents based on behavioral and long-term objectives
Financial markets also serve as an experimental testbed for our broader work on AI reasoning, because decisions have clearly defined consequences and can be evaluated continuously in changing environments.
Connecting the Three Areas
Although the human brain, AI systems, and financial markets appear very different, they lead to many of the same statistical questions:
By studying these questions across biological, artificial, and financial systems, our goal is to develop statistical and computational principles for learning and decision-making under uncertainty.