Identifying Multiple Personalities in Large Language Models with External Evaluation (2402.14805v1)

Abstract: As LLMs are integrated with human daily applications rapidly, many societal and ethical concerns are raised regarding the behavior of LLMs. One of the ways to comprehend LLMs' behavior is to analyze their personalities. Many recent studies quantify LLMs' personalities using self-assessment tests that are created for humans. Yet many critiques question the applicability and reliability of these self-assessment tests when applied to LLMs. In this paper, we investigate LLM personalities using an alternate personality measurement method, which we refer to as the external evaluation method, where instead of prompting LLMs with multiple-choice questions in the Likert scale, we evaluate LLMs' personalities by analyzing their responses toward open-ended situational questions using an external machine learning model. We first fine-tuned a Llama2-7B model as the MBTI personality predictor that outperforms the state-of-the-art models as the tool to analyze LLMs' responses. Then, we prompt the LLMs with situational questions and ask them to generate Twitter posts and comments, respectively, in order to assess their personalities when playing two different roles. Using the external personality evaluation method, we identify that the obtained personality types for LLMs are significantly different when generating posts versus comments, whereas humans show a consistent personality profile in these two different situations. This shows that LLMs can exhibit different personalities based on different scenarios, thus highlighting a fundamental difference between personality in LLMs and humans. With our work, we call for a re-evaluation of personality definition and measurement in LLMs.

• The paper introduces a novel external evaluation method to assess LLM personality traits by predicting MBTI types from generated text responses.
• It demonstrates that LLMs exhibit inconsistent personality profiles across roles, with notable differences between posts and comments.
• The methodology leverages a fine-tuned Llama2-7B model achieving over 80% MBTI classification accuracy, outperforming traditional models.

LLM Personality Analysis via External Evaluation

The paper "Identifying Multiple Personalities in LLMs with External Evaluation" (2402.14805) introduces a novel methodology for assessing the personality traits of LLMs. This approach diverges from traditional self-assessment methods, which have been shown to be unreliable due to the sensitivity of LLMs to prompt variations and question ordering. Instead, the authors propose an external evaluation method, leveraging a fine-tuned Llama2-7B model to predict the MBTI personality type based on the LLM's generated text in response to open-ended situational questions. The paper's findings reveal a significant inconsistency in the personality traits exhibited by LLMs when generating posts versus comments, challenging the notion that LLMs possess a stable, enduring personality akin to that of humans.

Methodology

The methodology comprises three key stages, as illustrated in (Figure 1):

Figure 1: This flowchart summarizes the steps in the methodology, from fine-tuning the Llama2-7B model to evaluating LLM personalities.

First, a Llama2-7B model is fine-tuned on a publicly available MBTI dataset to create a personality detection model. This fine-tuned model significantly outperforms existing state-of-the-art models in MBTI prediction accuracy, achieving over 80% accuracy in 16-class classification. The fine-tuning process employs LoRA, targeting the query and value layers. Second, the paper prompts LLMs (ChatGPT and Llama2 variants) to generate Twitter posts and comments based on summarized news events and existing tweets, respectively. This stage aims to elicit responses from LLMs in different roles, simulating real-world social media interactions. The prompts are designed to be simple to avoid influencing the LLMs' behavior. Finally, the generated text is fed into the fine-tuned personality detection model to obtain MBTI personality predictions. To mitigate sampling bias, 100 sets of 50 generations are created for each LLM and role, and the distribution of predicted MBTI types is analyzed.

Key Findings and Analysis

The paper's most striking finding is the significant difference in personality distributions exhibited by LLMs when generating posts versus comments. For example, the Llama2-7B model is predicted to have an ESTJ personality type when generating posts but an INFP type when generating comments. These inconsistencies are further visualized in (Figure 2), which shows the MBTI distribution of ChatGPT and Llama2 models.

Figure 2: These pie charts show the MBTI distribution of ChatGPT and Llama2 models, using the external evaluation method, highlighting the personality differences between generating posts and comments.

To validate the external evaluation method, the authors conducted a similar analysis on human-generated posts and comments from eight celebrities. The results, shown in (Figure 3), demonstrate that the personality distributions for humans remain relatively consistent across the two roles, aligning with the psychological definition of personality as an enduring characteristic.

Figure 3: These pie charts illustrate the MBTI distribution of four celebrities, using the external evaluation method, showing consistency between personality assessments from posts and comments.

This validation experiment confirms that the observed inconsistencies in LLM personalities are not artifacts of the personality detection model but rather reflect a fundamental difference in how LLMs and humans exhibit personality traits.

Implications and Future Directions

The findings of this paper have significant implications for how LLM behavior is understood and evaluated. The authors challenge the direct application of human personality frameworks to LLMs, arguing that the definition of personality may need to be re-evaluated in the context of artificial intelligence. The paper highlights the need for caution when using LLMs in applications where consistent and predictable behavior is critical. Future research should focus on developing new methods for defining and measuring personality in LLMs, taking into account their unique characteristics and capabilities. Additionally, further investigation into the factors that influence LLM personality shifts, such as prompt engineering and training data, could provide valuable insights into controlling and shaping LLM behavior.

Conclusion

The paper "Identifying Multiple Personalities in LLMs with External Evaluation" (2402.14805) presents a compelling case for re-evaluating the concept of personality in LLMs. By employing an external evaluation method and comparing LLM behavior to that of humans, the authors demonstrate that LLMs exhibit inconsistent personality traits across different contexts, challenging the notion that they possess a stable, enduring personality. This work calls for a more nuanced understanding of LLM behavior and the development of new frameworks for evaluating their personality traits.