Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Simulating Family Conversations using LLMs: Demonstration of Parenting Styles

Frank Tian-fang Ye, Xiaozi Gao

TL;DR

The paper tackles the challenge of studying parenting styles without human participants by using LLM-driven AI–AI conversations across four parenting categories. It introduces a programmable framework with context-aware prompts, two language models (Mixtral-8x7b-instruct and GPT-4-turbo), and context variations to simulate controlled family dialogues. Key findings show that the parenting styles are generally reflected in the simulations, model capability influences content diversity, and context/history improves coherence, with few-shot prompting further enhancing parental responses. This approach offers a flexible, low-cost platform for psycholinguistic and developmental research, enabling systematic manipulation of personalities, topics, and linguistic styles, while highlighting limitations related to model quality and potential for repetition that can be mitigated via fine-tuning and targeted prompting.

Abstract

This study presents a framework for conducting psychological and linguistic research through simulated conversations using large language models (LLMs). The proposed methodology offers significant advantages, particularly for simulating human interactions involving potential unethical language or behaviors that would be impermissible in traditional experiments with human participants. As a demonstration, we employed LLMs to simulate family conversations across four parenting styles (authoritarian, authoritative, permissive, and uninvolved). In general, we observed that the characteristics of the four parenting styles were portrayed in the simulated conversations. Several strategies could be used to improve the simulation quality, such as including context awareness, employing a few-shot prompting approach or fine-tuning models to cater to specific simulation requirements. Overall, this study introduces a promising methodology for conducting psychological and linguistic research through simulated conversations, while acknowledging the current limitations and proposing potential solutions for future refinement and improvement.

Simulating Family Conversations using LLMs: Demonstration of Parenting Styles

TL;DR

The paper tackles the challenge of studying parenting styles without human participants by using LLM-driven AI–AI conversations across four parenting categories. It introduces a programmable framework with context-aware prompts, two language models (Mixtral-8x7b-instruct and GPT-4-turbo), and context variations to simulate controlled family dialogues. Key findings show that the parenting styles are generally reflected in the simulations, model capability influences content diversity, and context/history improves coherence, with few-shot prompting further enhancing parental responses. This approach offers a flexible, low-cost platform for psycholinguistic and developmental research, enabling systematic manipulation of personalities, topics, and linguistic styles, while highlighting limitations related to model quality and potential for repetition that can be mitigated via fine-tuning and targeted prompting.

Abstract

This study presents a framework for conducting psychological and linguistic research through simulated conversations using large language models (LLMs). The proposed methodology offers significant advantages, particularly for simulating human interactions involving potential unethical language or behaviors that would be impermissible in traditional experiments with human participants. As a demonstration, we employed LLMs to simulate family conversations across four parenting styles (authoritarian, authoritative, permissive, and uninvolved). In general, we observed that the characteristics of the four parenting styles were portrayed in the simulated conversations. Several strategies could be used to improve the simulation quality, such as including context awareness, employing a few-shot prompting approach or fine-tuning models to cater to specific simulation requirements. Overall, this study introduces a promising methodology for conducting psychological and linguistic research through simulated conversations, while acknowledging the current limitations and proposing potential solutions for future refinement and improvement.
Paper Structure (14 sections, 4 figures, 1 table)

This paper contains 14 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Simulation Methods
  3. Models
  4. Conversation Context Awareness
  5. Prompt Engineering
  6. Simulation Setup
  7. Data Availability
  8. Results
  9. The parenting styles were generally reflected in the simulated conversations.
  10. The more capable LLM generated more diverse content.
  11. Feeding previous conversations to the agents generated better conversations.
  12. Few-shot learning significantly enhanced the parent's conversations.
  13. Discussion
  14. Conclusion

Figures (4)

  • Figure 1: Simulation setup for conversations with and without context awareness.
  • Figure 2: Embeddings of each conversation exchange were stored and retrieved to increase the agent's context awareness
  • Figure 3: Fine-tuned models may be used for better personality portray
  • Figure 4: Word frequencies of generated conversations