Injecting Instructions via Prompts Due to the lack ofexisting therapy transcripts, implementing a corpus-basedapproach would require a costly process of translating the

(1) Corpus-Based learning approaches train the agent ona given set of data to predict the next dialog act. They relyon large datasets and/or human annotation, and are consid-ered as inferior in optimizing for the long-term goal of theconversation. Such approaches are extended with planningtechniques such as knowledge graphs (Yang et al. 2022) andstochastic processes (Wang, Lin, and Li 2023). However,they do not suffice for our requirement for explicit formu-lation and following of expert instructions.

Structured Data: NumPy's Structured Arrays