Large Language Models (LLMs) in artificial intelligence have undergone extensive training with data created by humans, equating to some 20.000 years of human reading on an 8-10 day. This vast amount of knowledge allows LLMs to display human-like personas that sound quite convincing in their responses. This is often referred to as a synthetic personality and work may be underway towards building our Synthetic Personality of Large Language Models.

The idea of a human personality is a distinctive mix of attributes, features, and mental processes that form our core social relations and inclinations, stemming from our common biological and environmental past experiences (some argue that hope, belief, fear, and other projections into the future also play a part in the shaping of a human personality but they are still based on past experiences).

The rise of synthetic personalities in LLMs has triggered extensive research to comprehend the potential unexpected outcomes of these enhanced skills. Cases have been noted where LLMs have generated violent, deceitful, and manipulative language during tests, sparking worries about the reliability of dialogues, explanations, and knowledge derived from these models.

As LLMs become the primary medium for human-computer interaction, comprehending the characteristics associated with personality traits in the language produced by these models becomes vital. It’s also critical to learn how to safely and effectively craft personality profiles produced by LLMs. Researchers have examined techniques like few-shot prompting to reduce the influence of adverse and intense personality traits in LLM outcomes. However, the task of scientifically and systematically measuring their personalities still poses a challenge.

Researchers from institutions including Google DeepMind, the University of Cambridge, Google Research, Keio University, and the University of California, Berkeley, have suggested thorough, validated psychometric methods to shape and define personality synthesis based on LLMs.

The team has designed a method to use psychometric tests to confirm the validity of personality depiction in LLM-generated content. They’ve introduced a novel strategy to simulate population variance in LLM responses using controlled prompting. This approach checks the statistical associations between personality and its external correlates in human social science data. They’ve also offered a method to shape personality that functions independently of LLM, resulting in noticeable changes at the trait level.

The researchers validated their method on LLMs of different sizes and training techniques in two natural interaction environments: MCQA and long-form text generation. The results show that, given certain prompting configurations, LLMs can consistently and accurately mimic personality in their outputs. The evidence of reliability and validity of LLM-simulated personalities is more substantial for larger, fine-tuned models. Furthermore, the personality in LLM outputs can be adjusted along preferred dimensions to resemble specific personality profiles.

The introduction of LLMs has transformed natural language processing by allowing the creation of fluent and contextually appropriate text. As LLMs increasingly power conversational agents, the synthesized personality integrated into these models becomes a focus of interest. The comprehensive approach the researchers have taken to administer validated psychometric tests and quantify, analyze, and shape personality traits displayed in the text produced by commonly-used LLMs presents new opportunities. However, it also highlights significant ethical considerations, particularly regarding the responsible use of LLMs.

Takeaways

• Large language models (LLMs) are known to generate text that exhibits human-like personality traits. However, the complex psychosocial aspects of personality have not been thoroughly studied in LLM research.
• To ensure that LLM-based interactions are safe and predictable, it is important to quantify and validate these personality traits. This study provides a detailed quantitative analysis of personality traits in text generated by LLMs, using validated psychometric tests.
• The findings confirm that synthetic personality levels can be reliably measured through LLM-simulated psychometric test responses and LLM-generated text. This is especially true for larger and more finely-tuned models.
• The researchers also introduce methods for molding LLM personalities along desired dimensions to mimic specific personality profiles. They discuss the ethical considerations of such engineering of LLM personalities.