Skip to main content Skip to navigation

Retrieval-augmented Generation to Improve Math Question-Answering: Trade-offs Between Groundedness and Human Preference

Project Overview

The document explores the use of generative AI in education, specifically focusing on a Retrieval-Augmented Generation (RAG) system designed to enhance math question-answering for middle-school students. By incorporating external knowledge sources, this innovative approach aims to improve the quality of responses produced by large language models (LLMs). The study assesses various prompt guidance conditions to determine their influence on human preferences for the AI-generated answers, highlighting the importance of striking a balance between groundedness—ensuring the responses are based on reliable information—and user preference to optimize educational interactions. The findings underscore the potential of generative AI to support learning by providing tailored, context-aware assistance that meets students' needs while fostering engagement and understanding in complex subjects like math. Overall, this research demonstrates the effectiveness of integrating AI into educational frameworks, paving the way for more personalized and effective learning experiences.

Key Applications

Retrieval-augmented generation (RAG) system for math question-answering

Context: Middle-school math education, particularly for low-income students in Sierra Leone, Liberia, Ghana, and Rwanda.

Implementation: The RAG system uses a vetted open-source math textbook (OpenStax Prealgebra) to retrieve relevant information for answering student queries. Prompts were designed to influence the LLM's responses based on the retrieved content.

Outcomes: Improved response quality and student preferences for responses that are neither too grounded nor too vague. Preferences indicated that students favor responses that balance guidance and relevance.

Challenges: Responses can sometimes be incorrect or misaligned with the curriculum. There are trade-offs between groundedness and perceived usefulness of the responses.

Implementation Barriers

Technical Barrier

LLMs can generate incorrect answers or 'hallucinate' plausible yet factually incorrect information.

Proposed Solutions: Implement retrieval-augmented generation to draw from vetted educational materials and improve the correctness of the responses.

Implementation Barrier

Challenges in ensuring responses are well-aligned with specific educational contexts and curricula.

Proposed Solutions: Incorporate a variety of educational resources and maintain a focus on curriculum alignment during system design.

Project Team

Zachary Levonian

Researcher

Chenglu Li

Researcher

Wangda Zhu

Researcher

Anoushka Gade

Researcher

Owen Henkel

Researcher

Millie-Ellen Postle

Researcher

Wanli Xing

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Zachary Levonian, Chenglu Li, Wangda Zhu, Anoushka Gade, Owen Henkel, Millie-Ellen Postle, Wanli Xing

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies