# Experiences Build Characters: The Linguistic Origins and Functional Impact of LLM Personality

**Authors:** Xi Wang, Mengdie Zhuang, Jiqun Liu  
**arXiv:** [2603.06088](https://arxiv.org/abs/2603.06088)  
**Date:** 2026-03-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2603.06088v1/

## Key points

- Study by Xi Wang, Mengdie Zhuang, and Jiqun Liu on machine personality in Large Language Models (LLMs)
- Importance of diverse experiences in shaping machine personality and problem-solving
- Current LLM development prioritizes uniform performance benchmarks over varied traits
- Use of continued pre-training with domain-specific texts to mimic experience accumulation
- Adaptation of Big Five framework into Machine Personality Inventory (MPI) for quantifying model traits
- Identification of "Suppression Advantage" phenomenon where reduced social traits enhance reasoning performance
- Causal link between training data linguistics and lexical diversity within models
- Insights into how machine personality influences problem-solving capabilities in LLMs
- Roadmap for "Personality Engineering" to optimize model training based on personality traits

## Summary

In their study "Experiences Build Characters: The Linguistic Origins and Functional Impact of LLM Personality," Xi Wang, Mengdie Zhuang, and Jiqun Liu explore how diverse experiences shape machine personality and influence problem-solving in Large Language Models (LLMs). They highlight the importance of incorporating a variety of styles and personality traits in human problem-solving, yet current LLM development often prioritizes uniform performance benchmarks that favor specific behavioral tendencies like assertiveness. To address this gap, the researchers employ continued pre-training to expose models to domain-specific texts in an unsupervised manner, mimicking the accumulation of experience. By adapting the Big Five framework through the Machine Personality Inventory (MPI), they quantify the personality traits of different model variants and analyze how these traits relate to linguistic style and reasoning behavior. Interestingly, they identify a phenomenon called the "Suppression Advantage," where reduced social traits actually enhance complex reasoning performance in LLMs. Additionally, the research establishes a causal link between training data linguistics—such as imperative frequency—and lexical diversity within models. Overall, this study provides valuable insights into how machine personality is shaped by experiences and how it impacts problem-solving capabilities in LLMs. The findings offer a roadmap for "Personality Engineering," suggesting ways to optimize model training for improved performance based on personality traits.