Thinking Assistants: LLM-Based Conversational Assistants that Help Users Think By Asking rather than Answering

• Main Authors: Park, Soya, Kulkarni, Chinmay
• Format: Journal Article
• Language: English
• Published: 10.12.2023
• Subjects: Computer Science - Human-Computer Interaction
• DOI: 10.48550/arxiv.2312.06024

We introduce the concept of "thinking assistants", an approach that encourages users to engage in deep reflection and critical thinking through brainstorming and thought-provoking queries. We instantiate one such thinking assistant, Gradschool.chat, as a virtual assistant tailored to assist prospective graduate students. We posit that thinking assistants are particularly relevant to situations like applying to graduate school, a phase often characterized by the challenges of academic preparation and the development of a unique research identity. In such situations, students often lack direct mentorship from professors, or may feel hesitant to approach faculty with their queries, making thinking assistants particularly useful. Leveraging a Large Language Model (LLM), Gradschool.chat is a demonstration system built as a thinking assistant for working with specific professors in the field of human-computer interaction (HCI). It was designed through training on information specific to these professors and a validation processes in collaboration with these academics. This technical report delineates the system's architecture and offers a preliminary analysis of our deployment study. Additionally, this report covers the spectrum of questions posed to our chatbots by users. The system recorded 223 conversations, with participants responding positively to approximately 65% of responses. Our findings indicate that users who discuss and brainstorm their research interests with Gradschool.chat engage more deeply, often interacting with the chatbot twice as long compared to those who only pose questions about professors.