In a recent study, math problems in Carnegie Learning’s MATHia adaptive learning software were rewritten by human authors and AI to improve clarity. Findings showed that readers spent less time reading rewritten human content and achieved higher mastery than did readers who read the original content. The team conducting the study also used GPT-4 to rewrite the same set of math word problems with the same guidelines that the human authors used. comparing zero-shot, few-shot and chain-of-thought prompting strategies. Overall, report analysis of human-written, original and GTP-written problems showed that GTP rewrites have the most optimal readability, lexical diversity and cohesion scores, though used more low frequency words. Carnegie Learning plans to present their outputs at randomized field trials in MATHia.
Retrieval-augmented Generation to Improve Math Question-Answering: Trade-offs Between Groundedness and Human Preference
Interactive question-answering (QA) with tutors has shown to be an effective way for middle school math students to learn. While not all students have access to a tutor, large language models make it possible to automate portions of the tutoring process–including interactive QA to support students’ conceptual discussion of mathematical concepts. Some have questioned how LLM responses can be better aligned with a school’s curriculum. In this paper, Levonian and colleagues explore how retrieval-augmented generation (RAG) can help improve response quality by incorporating textbook information and other educational resources, while also identifying trade-offs of using RAG.