# Personality Traits in Large Language Models

**Authors:** Mustafa Safdari, Greg Serapio-García, Clément Crepy, Stephen Fitz, Peter Romero, Luning Sun, Marwa Abdulhai, Aleksandra Faust, Maja Matarić  
**arXiv:** [2307.00184](https://arxiv.org/abs/2307.00184)  
**Date:** 2023-07-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.00184v1/

## Key points

- Impact of large language models (LLMs) on natural language processing and conversational agents
- Importance of personality in communication
- Method for administering psychometric tests to quantify and analyze personality traits exhibited in LLM-generated text
- Simulated personality in some LLM outputs is reliable and valid, especially for larger models with instruction fine-tuning
- Personality in LLM outputs can be shaped to mimic specific profiles
- Potential applications and ethical implications of the measurement and shaping framework
- Overview of LLMs as large-capacity machine learning models for generating coherent and contextually relevant text
- Learning process of LLMs from various datasets including web examples, instruction datasets, and preference fine-tuning
- Previous attempts to probe personality traits in LLMs lacking rigor compared to standard quantitative social sciences practices
- Psychometrics-grounded construct validity when administering personality inventories to LLMs
- Evaluation and shaping methodology for personalities in LLM outputs using validated psychometric tests and prompt structures
- Reliable and valid simulated personalities in some LLM outputs under specific prompting configurations, stronger evidence for larger models with instruction fine-tuning
- Shaping personality dimensions to mimic specific profiles
- Potential applications and ethical considerations of the measurement and shaping framework
- Responsible use of LLMs emphasized throughout the paper

## Summary

This paper discusses the impact of large language models (LLMs) on natural language processing and conversational agents. The authors highlight the importance of personality in communication and present a comprehensive method for administering psychometric tests to quantify and analyze personality traits exhibited in text generated by LLMs. They find that the simulated personality in some LLM outputs is reliable and valid, especially for larger models that have undergone instruction fine-tuning. Additionally, they demonstrate that personality in LLM outputs can be shaped to mimic specific profiles. The authors also discuss potential applications and ethical implications of their measurement and shaping framework, emphasizing responsible use of LLMs. In the introduction, the authors provide an overview of LLMs as large-capacity machine learning models that generate coherent and contextually relevant text. They explain how LLMs have revolutionized natural language processing and meet key requirements for human-like conversation. The authors mention that LLMs learn from large datasets of text from various sources including web examples "in context," instruction datasets, and preference fine-tuning. The related work section discusses previous attempts to probe personality traits in LLMs. While some studies suggest that certain models exhibit dark personality patterns or administer personality inventories to LLMs, the rigor in evaluating these results has been limited compared to standard quantitative social sciences practices. In contrast, this paper aims to establish psychometrics-grounded construct validity when administering personality inventories to LLMs. The methodology section provides details on the evaluation and shaping of personalities in LLM outputs. The authors describe their approach for administering validated psychometric tests to measure personality traits exhibited in generated text. They also outline prompt structures used for evaluation purposes. The findings section presents the results of their study. The authors report that simulated personalities in some LLM outputs are reliable and valid under specific prompting configurations with stronger evidence for larger models that have undergone instruction fine-tuning. Furthermore, they demonstrate that personality in LLM outputs can be shaped along desired dimensions to mimic specific profiles. The implications section discusses the potential applications and ethical considerations of the measurement and shaping framework developed by the authors. They emphasize responsible use of LLMs and highlight the need for further research on its ethical implications when using them to generate text with specific personalities. In conclusion, this paper presents a comprehensive method for administering psychometric tests to quantify and analyze personality traits exhibited in text generated by large language models (LLMs). The authors demonstrate reliability, validity, as well as shapeability of simulated personalities in such outputs while discussing potential applications as well as ethical implications associated with their framework - emphasizing responsible use thereof throughout their work..