Publication Date

Fall 2025

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Christopher Pollett

Second Advisor

Robert Chun

Third Advisor

Sayma Akther

Keywords

Visual Cortex, AI Agents, Topology Manager, Phi-3.5-mini-instruct

Abstract

Agent based systems can simulate a variety of complex environments, social behavior and
cognitive processes. Since the heart of these agents, Large Language Models (LLMs) are trained
on data generated by humans, they offer a route to seamless human cognition modelling. In this
work, we build simulations of the visual cortex using the dorsal-ventral model of cognitive vision
and further extend it to model V1–V6 visual processing subsystems, that perform functions like
contour detection, binocular disparity, motion perception and color perception that support visual
processing in the human brain. To implement this, we developed a framework, Topology Manger
to enable deployment and management of LLM agents with customizable interaction patterns,
memory management, and support for various LLM backends. We also finetuned a Phi-3.5-mini-
instruct model on the ReClor logical reasoning dataset, achieving a prediction accuracy of 89%,
which will later be used to maintain coherent reasoning across the simulated brain regions. Using
these components, we construct a system that predicts human visual attention given an input
image, achieving a precision score of 0.6 on a standard visual saliency dataset. This work
demonstrates the potential of multi-agent LLM systems in modeling brain processes and
contributes toward building cognitively grounded AI architectures.

Download

Available for download on Friday, December 18, 2026

Share

COinS