In this research, we investigate the use of Large Language Models (LLMs) to support elementary math education. Specifically, we examined various methods of prompting an LLM, in this case GPT-3, to provide multi-step explanations for various math concepts covered in the elementary math curriculum from Khan Academy. We evaluated the effectiveness of these different prompting methods and used this information to develop MathyAI, a conversational virtual assistant for elementary math. Finally, we conducted initial user studies to assess the potential of MathyAI and identify areas for improvement. While our work relates to the use of LLMs for solving math problems, our primary focus was on the educational potential of these models.
This project took a deep dive into the field of text summarization and specifically focused on news text. We compared and analyzed the difference between a frequency-based model and a Seq2Seq model quantitatively and qualitatively using ROUGE variants as our evaluation metrics. As expected, while the neural network-based Seq2Seq model performed better than the conventional frequency-based in news summarization, the frequency-based model has its merits. Our project would enrich the context of natural language understanding(NLU) by analyzing the recent news dataset and suggesting future researchers who are interested in studying news summarization try using a combination of an abstractive model and an extractive model.
Our project was interested in learning infant self-play behavior, specifically how they learn about the physical dynamics of objects and environments around them through interaction and observation. We took inspiration from Model Predictive Control (MPC) and designed the model predictive curiosity (MPCu) paradigm. Then we set up NVIDIA’s IsaacGym physical simulation environment and Box2D environment. We generated 50000 training scenarios of a force being applied to a circle adjacent to a tower in a Box2D environment and trained a dynamics model to predict forward motion of circle, and a curiosity model to predict the loss in the dynamics model. We believe our framework MPCu is capable of directly optimizing for high curiosity action values and enriching forces that cause multi-object interactions.
Our project was interested in studying what factors make an answer on Stack Overflow popular. We used SQL to explore the stackoverflow dataset from BigQuery Public Dataset, analyzed the popularity of stack overflow’s answers based on its votes and other characteristics and then visualized the correlations. We found answer count, comment count and score to be the most influential features of the number of votes. Based on our data explorations, we fitted two models: logistic regression and boosted tree classifier to our data and both models resembled fair performance, boosted tree classifier provided predictions with higher quality in terms of accuracy, precision, and ROC-AUC across all three subsets of data: training, validation and testing.
We conducted a secondary analysis to examine the usage of two interactive e-book features, Multimedia Dictionary and Story Character Statements, and their association with children’s learning and reading motivation outcomes. We found that Story Character Statement hotspots with content congruent to the storyline are associated with lower story comprehension post-test scores and Multimedia Dictionary with content congruent to the story vocabulary but incongruent to the story associated with higher comprehension and retelling scores. These results challenged previous conclusions in the literature that the content of e-book design features has to align with the storyline to support children’s learning. In summary, our findings provided critical implications for future research on e-book design and learning.