Gabriel Thompson

Gabriel Thompson

Biography

I am Gabriel Thompson, a rising senior at North Carolina State University studying physics and electrical engineering. As of late, I have been interested in machine learning and AI. Specifically, I want to i) further our scientific understanding of machine learning and ii)use this understanding to design better models and promote safe and responsible usage ofAI.My first academic love was physics – the idea that you could precisely understand the world through mathematics was absolutely fascinating. I aim to bring a similar, first-principles mindset to projects that I am involved in. I have previously done research for a science communication initiative at NC State, interned as a data scientist in supply chain, and authored a paper in privacy-preserving machine learning. Aside from work, I play music in my free time. I also ran a marathon, love to exercise, and plan to bike across Europe

Abstract

Teaching Language Models to Ask Informative Questions

As the adoption of artificial intelligence (AI) continues to grow, researchers have focused on the question of how to align AI models to human preferences. While current work like RLHF addresses this alignment in language models, it typically overlooks the question of how human preferences can be elicited in the first place. Previous work explicitly instructs the large language model (LLM) to elicit the user’s preferences. However, LLMs are known to be poor sequential decision-makers. To alleviate this issue, we propose an approach that teaches the LLM to ask informative questions. Specifically, we use expected information gain to fine-tune the model on Bayesian-optimal questions, teaching them to reason probabilistically. We find that the LLM is able to learn from this Bayesian signal; performing better on not only the fine-tuning task, but also generalizing beyond the context in which it was trained.