Author ORCID Identifier

https://orcid.org/0009-0007-3926-255X

Date of Award

Spring 4-8-2025

Document Type

Thesis (Undergraduate)

Department

Computer Science

First Advisor

Hae Won Park

Second Advisor

Pedro Colon-Hernandez

Third Advisor

Lorie Loeb

Abstract

Recent advancements in large language models (LLMs) have demonstrated their ability to leverage its immense world knowledge to excel in traditional reinforcement learning tasks. Building on these capabilities, we introduce a framework that uses LLMs to learn from human behavior video data and generate insights that serve as guidelines for predicting how individuals are likely to act. Our framework focuses on parent- child dialogic reading, with an emphasis on understanding each parent’s unique parenting strategies. The system identifies meaningful interaction episodes within analyzed video data. The learning component of the framework utilizes a long-term memory system, which stores and retrieves relevant insights from past interactions based on their similarity to the current context. We introduce a trajectory-based approach that treats reflection as a reward for the model, accounting for instances where the optimal choice may not be available in the training data. The resulting model generates personalized parenting strategies by applying stored insights to new situations. It continuously learns from the parent’s dialogic behaviors, refines its understanding through accumulated insights, and generates strategies that align with the parent’s style.

Share

COinS