Prompt Engineering an Informational Chatbot for Educating about Mental Health: Utilizing a Multi-Agent Approach for Enhanced Compliance with Prompt Instruction (Preprint)

JMIR AI, Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 9, 2024

People with schizophrenia often present cognitive impairments that may hinder their ability to learn about condition. Education platforms powered by Large Language Models (LLMs) have the potential improve accessibility of mental health information. However, black-box nature LLMs raises ethical and safety concerns regarding controllability over chatbots. In particular, prompt-engineered chatbots drift from intended role as conversation progresses become more prone hallucinations. To develop evaluate a Critical Analysis Filter (CAF) system ensures an LLM-powered chatbot reliably complies predefined its instructions scope while delivering validated For proof-of-concept, we educational GPT-4 can dynamically access information manual written for people caregivers. CAF, team LLM agents are used critically analyze refine chatbot's responses deliver real-time feedback chatbot. assess CAF re-establish adherence instructions, generate three conversations (by conversing disabled) wherein starts towards various unintended roles. We use these checkpoint initialize automated between adversarial designed entice it Conversations were repeatedly sampled enabled disabled respectively. Three human raters independently rated each response according criteria developed measure integrity; specifically, transparency (such admitting when statement lacks explicit support scripted sources) tendency faithfully convey in manual. total, 36 (3 different conversations, 3 per checkpoint, 4 queries conversation) compliance Activating resulted score was considered acceptable (≥2) 67.0% responses, compared only 8.7% deactivated. Although rigorous testing realistic scenarios is needed, our results suggest self-reflection mechanisms could enable be effectively safely platforms. This approach harnesses flexibility constraining appropriate accurate interactions.

Language: Английский

Prompt Engineering an Informational Chatbot for Education on Mental Health Using a Multiagent Approach for Enhanced Compliance With Prompt Instructions: Algorithm Development and Validation (Preprint)

Published: Dec. 9, 2024

People with schizophrenia often present cognitive impairments that may hinder their ability to learn about condition. Education platforms powered by large language models (LLMs) have the potential improve accessibility of mental health information. However, black-box nature LLMs raises ethical and safety concerns regarding controllability chatbots. In particular, prompt-engineered chatbots drift from intended role as conversation progresses become more prone hallucinations. This study aimed develop evaluate a critical analysis filter (CAF) system ensures an LLM-powered chatbot reliably complies its predefined instructions scope while delivering validated For proof concept, we prompt engineered educational for GPT-4 could dynamically access information manual written people caregivers. CAF, team LLM agents was used critically analyze refine chatbot’s responses deliver real-time feedback chatbot. To assess CAF re-establish adherence instructions, generated 3 conversations (by conversing disabled) wherein started toward various unintended roles. We these checkpoint initialize automated between adversarial designed entice it Conversations were repeatedly sampled enabled disabled. total, human raters independently rated each response according criteria developed measure integrity, specifically, transparency (such admitting when statement lacked explicit support scripted sources) tendency faithfully convey in manual. 36 (3 different conversations, per checkpoint, 4 queries conversation) compliance Activating resulted score considered acceptable (≥2) 81% (7/36) responses, compared only 8.3% (3/36) deactivated. Although rigorous testing realistic scenarios is needed, our results suggest self-reflection mechanisms enable be effectively safely platforms. approach harnesses flexibility constraining appropriate accurate interactions.

Language: Английский