Characterizing Situational Control Assessment Mechanisms for Emotion Prediction Tasks in a Large Language Model

open_access
Open Access
Author:
Craig, Lucas
Area of Honors:
Computer Science
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
James Z Wang, Thesis Supervisor
Sencun Zhu, Thesis Honors Advisor
Shagufta Mehnaz, Faculty Reader
Keywords:
large language model
emotion
cognition
explainable ai
mechanistic interpretability
theory of mind
ai
machine learning
Abstract:
As Large language models (LLMs) have become better at imitating and understanding human behaviors, their ability to predict human emotional responses to situations has improved. Prior work has shown that these predictions are influenced by cognitive appraisal dimensions, of the kind proposed by the appraisal theory of emotion in psychology. More recent studies have identified the general location of the appraisal within LLMs, but our understanding of how LLMs perform appraisals remains fairly limited. In this thesis, I take a step towards bridging this gap by studying a small but capable LLM: Qwen3-4B. Using a set of minimally-differing prompts, along with various tools from the “circuit analysis” sub-field of AI interpretability, I attempt to map out the mechanisms responsible for using human-controllable vs. situational appraisals of negative events to inform emotion predictions. I identify a circuit that modulates the anger-vs-sadness distinction on a negative-valence emotion prediction task; testing the circuit across a broad set of agent types reveals that it does not cleanly implement human controllability as defined by appraisal theory. Instead, it appears to detect something like autonomous animate agency, modulated by intentionality modifiers (e.g. "deliberately," "accidentally"). These results suggest that Qwen3-4B has learned a situational control-adjacent appraisal dimension, one which also bundles together distinctions that psychological appraisal theory keeps separate (e.g. human vs. animal agency, intentionality).