Date of Award

Spring 5-8-2025

Document Type

Thesis (Master's)

Department or Program

Computer Science

First Advisor

Soroush Vosoughi

Second Advisor

Lorie Loeb

Third Advisor

John Bell

Abstract

This thesis presents a method for teaching large language models (LLMs) to perform embodied multi-task planning using simulation data and a lightweight training technique called proxy-tuning. Traditional LLMs are strong in language understanding but struggle with tasks that require reasoning and planning in real-world environments. To address this, we construct a dataset based on VirtualHome, a simulator that models household activities using input commands. We extract individual tasks from the simulator and combine them into multi-task plans, optimizing for coherence and efficiency. These plans are formatted as Q&A pairs to train models to generate step-by-step instructions that complete multiple tasks at once. Instead of directly fine-tuning large models, which is computationally expensive, we first fine-tune smaller models and then transfer their learned behavior to a larger model using proxy-tuning. This technique adjusts the output of the large model based on the differences between a fine-tuned and an unfine-tuned smaller model, without changing the large model’s parameters.

Share

COinS