Author ORCID Identifier

https://orcid.org/0000-0001-6615-2224

Date of Award

Spring 5-13-2026

Document Type

Thesis (Ph.D.)

Department or Program

Psychological & Brain Sciences

First Advisor

Emily S. Finn

Abstract

Human behavior and cognition are strikingly variable: people differ from one another in their preferences and abilities, and even from themselves across situations. Yet this diversity arises from common neural machinery shaped by the complex environments humans inhabit. A central objective of cognitive neuroscience is to understand how this varied experience emerges from interactions between brains, agents, and environments. In this dissertation, I argue that comparing behavior and neural computation across humans, artificial systems, and contexts is essential for understanding the flexibility of human cognition. Across three chapters, I use this comparative approach to examine how the rich, multimodal contexts of everyday experience shape human language and cognition.

In Chapter 2, I examine how individuals differ in their representations of language, and whether similar properties shape the representational spaces of computational models. Using brain activity recorded while participants listened to naturalistic stories, I show that the concrete-abstract axis –– a proxy for sensory grounding –– supports reliable representations of language that are unique to the individual. This was driven by the stability of concrete words, suggesting that sensory grounding provides a foundation for how individuals represent language. In Chapter 3, I evaluate how sensory context shapes human language predictions and their alignment with large language models (LLMs). I collected a behavioral dataset in which participants predicted upcoming words while reading, listening to, or watching real-world narratives. Humans consistently outperformed LLMs, and this advantage grew with greater access to sensory context. Through \textit{in-silico} experiments, I show that sensory context provides unique information for prediction, improving performance and accelerating language acquisition within LLMs. In Chapter 4, I investigate how experimental contexts shape the information encoded in human brain activity. By isolating shared, stimulus-driven activity across participants, I show that movie-watching drives stimulus-related responses further into higher-order association cortices than traditional tasks. While both paradigms relied on overlapping representations, only those derived from movies generalized to predict task-related brain activity, suggesting that naturalistic stimuli more efficiently sample the representational spaces underlying human cognition. Together, these studies demonstrate the power of comparative approaches as a framework for identifying the components that undergird human behavior and neural computation.

Download

Share

COinS