Name: Huadong Xiong

Ph.D. Dissertation Proposal Meeting

Time: Wednesday, February 18, 2026, 12:00 PM – 1:00 PM

Location: Zoom – https://gatech.zoom.us/my/huadong.xiong?pwd=u8JPjJbQwVLXhhf9RPItRJS7dcFEw4.1

Advisor:
Dr. Robert Wilson (Georgia Institute of Technology)

Dissertation Committee Members:
Dr. Dobromir Rahnev (Georgia Institute of Technology)
Dr. Apurva Ratan Murty (Georgia Institute of Technology)
Dr. Audrey Sederberg (Georgia Institute of Technology)
Dr. Marcelo Mattar (New York University)

Title: An Online Learning Perspective on Cognition

Abstract:

Both biological and artificial intelligence systems possess the capacity to adapt to novel task demands without structural retraining. In Large Language Models (LLMs), this capability manifests as in-context learning (ICL), where behavior is modified solely through inference on recent examples. Similarly, humans and animals exhibit "learning to learn," dynamically adjusting their integration and exploration strategies in response to environmental feedback. Despite these parallels, a unified computational framework governing how such strategies are updated online remains elusive.

In this PhD project, I propose an online learning framework to bridge the gap between human cognitive adaptation and LLM in-context learning. I hypothesize that cognition can be modeled as a continuous optimization process over an internal strategy space. By synthesizing behavioral modeling, gradient-based meta-learning, and geometric representational analysis, I will characterize strategy adaptation as an error-driven, resource-constrained process.

This research aims to provide a mechanistic account of adaptive intelligence, explicitly connecting the temporal dynamics of human perceptual and reinforcement learning with the representational reorganization observed in LLMs.