Extended Data Table 5 Associations between instructed and real-world task levels of synthetic personality for the largest model of each tested LLM family, presented as Spearman’s rank correlation coefficients (ρ)

From: A psychometric framework for evaluating and shaping personality traits in large language models

  1. Prompted (ordinal) levels of personality strongly relate to personality levels observed in synthetically-generated social media status updates for all Big Five traits, except openness—which is moderately correlated with target levels for Flan-PaLM 540B—demonstrating that LLM personality can be verifiably shaped in generative tasks for sufficiently-powerful models. All correlations are statistically significant at p < 0.0001 (two-sided values computed using Student’s t-distribution); n = 450 observations per targeted trait.
Back to article page