Generative Agents to Support Students Learning Progress

Abstract

Ongoing assessments in a course are crucial for tracking student performance and progress. However, generating and evaluating tests for each lesson and student can be time-consuming. Existing models for generating and evaluating question-answer pairs have had limited success. In recent years, large language models (LLMs) have become available as a service, offering more intelligent answering and evaluation capabilities. This research aims to leverage LLMs for generating questions, model answers, and evaluations while providing valuable feedback to students and decentralizing the dependency on faculty. Our approach is based on the development and interaction of advanced AI-powered generative agents, built on large language models like ChatGPT, GPT-4, and Vicuna, and designed to emulate human activities such as information abstraction, context refinement, and query rating. These agents interact autonomously in a network, employing techniques like Zero-Shot and Few-Shot Prompting to generate responses and adapt to various roles and contexts. The setup includes three key agents for question generation, refinement, and quality assurance, which leverage text vectorization, document selection and filtering, and cutting-edge language models to generate, refine, and evaluate questions and answers based on specific learning objectives. In conclusion, this paper demonstrates the versatility of LLMs for various learning tasks, including question generation, model answer generation, and evaluation, all while providing personalized feedback to students. By identifying and addressing knowledge gaps, LLMs can support continuous assessment and help students improve their understanding before semester exams.

Mehr zum Titel