# Roleplay-doh: Enabling Domain-Experts to Create LLM-simulated Patients via Eliciting and Adhering to Principles

**Authors:** Ryan Louie, Ananjan Nandi, William Fang, Cheng Chang, Emma Brunskill, Diyi Yang  
**arXiv:** [2407.00870](https://arxiv.org/abs/2407.00870)  
**Date:** 2024-07-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.00870v2/

## Key points

- Large Language Models (LLMs) are being used to simulate realistic social scenarios for improving social skills.
- Simulating sensitive interactions related to mental health presents challenges due to privacy concerns and the need for expert feedback.
- Roleplay-doh is an innovative approach developed by a team of researchers led by Ryan Louie, Ananjan Nandi, William Fang, Cheng Chang, Emma Brunskill, and Diyi Yang.
- Roleplay-doh involves a collaboration pipeline between humans and LLMs where domain experts provide qualitative feedback translated into natural language rules governing roleplay prompts.
- Specifically applied for senior mental health supporters creating customized AI patients as simulated practice partners for novice counselors.
- Researchers identified issues with GPT-4 simulations not aligning with expert-defined principles and introduced a principle-adherence prompting pipeline resulting in improved response quality.
- A user study involving 25 counseling experts showed that the Roleplay-doh pipeline makes it easier and more efficient to create AI patients resembling real patients according to creators and third-party counselors.

## Summary

In recent research, there has been a growing interest in using Large Language Models (LLMs) to simulate realistic social scenarios. This is particularly helpful for individuals looking to improve their social skills. However, simulating sensitive interactions related to mental health poses significant challenges that need to be addressed. Privacy concerns often limit access to the necessary data and obtaining expert feedback can be time-consuming. To tackle these issues, a team of researchers led by Ryan Louie, Ananjan Nandi, William Fang, Cheng Chang, Emma Brunskill, and Diyi Yang have developed Roleplay-doh - an innovative approach that involves a collaboration pipeline between humans and LLMs. This pipeline enables domain experts to provide qualitative feedback which is then translated into a set of principles or natural language rules governing the roleplay prompts generated by the LLM. Specifically applied for senior mental health supporters creating customized AI patients as simulated practice partners for novice counselors. Through this process, the researchers identified issues with GPT-4 simulations not aligning with expert-defined principles. To address this discrepancy, they introduced a novel principle-adherence prompting pipeline resulting in a 30% improvement in response quality and principle adherence for the downstream task. A user study involving 25 counseling experts was conducted to evaluate the effectiveness of the Roleplay-doh pipeline. The results demonstrated that this approach makes it easier and more efficient to create AI patients that closely resemble real patients according to both creators and third-party counselors. For those interested in exploring further details about Roleplay-doh and accessing code and data related to the project, the researchers have provided a project website at https://roleplay-doh.github.io/.