Chatting with GPT: Pragmatic-Pedagogic Dialogue Games

Abstract

The human-robot interaction community has shown growing interest in Large Language Model (LLM) powered agents like ChatGPT and LLaMA for their advanced communication skills. However, these fine-tuned LLMs often lack strategic conversation abilities. They tend to provide lengthy, verbose responses without considering their usefulness to humans, and they may not effectively gather information from the humans they assist. Generally, LLMs don't match human conversational skills, a shortfall possibly due to their training. Typically, LLMs are fine-tuned through a one-shot process based on human or AI-evaluated prompt rankings, which might not support sustained, helpful human interaction. This thesis investigates the use of game-theoretic tools for more effective conversational LLM fine-tuning, with a specific focus on the issues of developing computationally tractable algorithms with appealing theoretical properties (eg, convergence) to enable LLMs to handle value alignment even in the face of uncertainty.