AAI_2025_Capstone_Chronicles_Combined

17

Figure 4 Reconstruction loss curve during autoencoder training.

interpretations appear to be largely coherent and distinct, although there is some overlap between the “Extended Practice Across Topics” and “Diverse Problem-Solving Approaches” clusters. An example conversation assigned to the “Step-By-Step Guided Practice” cluster follows: User: I need 10 practice questions! Assistant: Absolutely! What area of math or science are you interested in practicing? User: precalculus Assistant: Great! Let’s get started with some precalculus practice. Here’s your first question: Question 1: If f ( x )= 3 x 2 + 2 x − 1, what is f ( − 2 ) ? Take as much time as you need and tell me your answer when you’re ready. User: 7 In human evaluation, 55% of the conversations matching the generated cluster description. This outperforms random chance, but is far below the consistency needed to create clusters that would provide valuable insights in practice. The k-means model outperformed the DEC model by

241