Many researchers are working to address the worldwide mental health crisis by developing therapeutic technologies that increase the accessibility of care, including leveraging large language model (LLM) capabilities in chatbots and socially assistive robots (SARs) used for therapeutic applications. Yet, the effects of these technologies over time remain unexplored. In this study, we use a factorial design to assess the impact of embodiment and time spent engaging in therapeutic exercises on participant disclosures. We assessed transcripts gathered from a two-week study in which 26 university student participants completed daily interactive Cognitive Behavioral Therapy (CBT) exercises in their residences using either an LLM-powered SAR or a disembodied chatbot. We evaluated the levels of active engagement and high intimacy of their disclosures (opinions, judgments, and emotions) during each session and over time. Our findings show significant interactions between time and embodiment for both outcome measures: participant engagement and intimacy increased over time in the physical robot condition, while both measures decreased in the chatbot condition.

Researchers have been interested in developing AI tools to help students learn various mathematical subjects. One challenging set of tasks for school students is learning to solve math word problems. We explore how recent advances in natural language processing, specifically the rise of powerful transformer based models, can be applied to help math learners with such problems. Concretely, we evaluate the use of GPT-3, a 1.75B parameter transformer model recently released by OpenAI, for three related challenges pertaining to math word problems corresponding to systems of two linear equations. The three challenges are classifying word problems, extracting equations from word problems, and generating word problems. For the first challenge, we define a set of problem classes and find that GPT-3 has generally very high accuracy in classifying word problems (80%-100%), for all but one of these classes. For the second challenge, we find the accuracy for extracting equations improves with number of examples provided to the model, ranging from an accuracy of 31% for zero-shot learning to about 69% using 3-shot learning, which is further improved to a high value of 80% with fine-tuning. For the third challenge, we find that GPT-3 is able to generate problems with accuracy ranging from 33% to 93%, depending on the problem type.