The improvement of Large Language Models (LLMs) signifies great opportunities for enhancing AI-assisted math teaching and learning. Wanli and Neil’s teams have worked to accelerate the pre-training, fine-tuning, and evaluation of LLMs for math education. By leveraging the University of Florida’s supercomputer and Microsoft’s DeepSpeed while utilizing ASSISTments’ datasets, the teams came to important findings about LLMs – including how super-sized LLMs such as GPT-3.5 may not always be the best option for specific downstream tasks. Overall, the preliminary results showed an enhancement of multimodal LLM models in accurate scoring and alignment with teacher feedback.

Because artificial intelligence or machine learning models are trained on data from different contexts, they are also susceptible to biases introduced through the selection and processing of datasets. Ensuring that learned models do not introduce or amplify systematic biases against underrepresented or historically marginalized groups is critical. This guide from the Fairness Hub serves as a starting point for more contextualized examinations of bias and fairness by providing (1) a simple metric for measuring bias, (2) an overview of bias mitigation strategies, (3) an overview of bias analysis and mitigation toolkits, and (4) a quick demonstration of how to measure and mitigate bias with the help of a toolkit.

Born out of Carnegie Mellon University 25 years ago, Carnegie Learning has been on the cusp of cutting-edge educational technology for nearly a quarter-century, fine-tuning products to help students learn mathematics best via years of data analytics and software improvements.