Author ORCID Identifier

https://orcid.org/0009-0008-5167-9048

Date of Award

Spring 6-3-2025

Document Type

Thesis (Undergraduate)

Department

Computer Science

First Advisor

Soroush Vosoughi

Abstract

This thesis investigates the capabilities of large language models (LLMs), specifically GPT4 and GPT4o, in appraising human emotional responses within strategic social scenarios. Building on the work of Houlihan et. al[7], we evaluate LLMs using a dataset of human emotion ratings from game-theoretic situations, later extending the experimental paradigm to include both text and multimodal (image and profession) inputs. Our methodology introduces novel prompting techniques for the experiment at hand, and we compare the performance of these techniques with expert perspectives to assess how different prompts influence model predictions. Results show that LLMs can approximate human emotional appraisals, with the first-person deixis yielding the closest alignment to empirical data. Comparative analysis between GPT4 and GPT4o reveals some differences, particularly in complex social contexts. We discuss the limitations of current LLMs, including their lack of explicit causal reasoning and challenges in generalizing across diverse scenarios. Finally, as an extension we propose a neuro-symbolic approach that integrates LLMs with Houlihan et al.’s computer appraisal model, aiming to combine the flexibility of neural networks with the interpretability of symbolic reasoning. This work advances the field of AI social cognition and highlights promising directions for developing emotionally intelligent, human-aligned artificial agents.

Share

COinS