Same Persona, Different Behaviors: Quantifying Persona Conditioning in Large Language Models

Abstract

Persona conditioning is increasingly used to guide the behavior of Large Language Models (LLMs), yet the extent to which models consistently and effectively adapt to assigned identities remains unclear. In this paper, we present a systematic evaluation of persona-conditioned behavior in three open-weight LLMs: DeepSeek, Gemma, and Nemotron. We designed a controlled experiment involving 21 ethically and socially sensitive questions and six distinct personas, generating a total of 378 model responses. Our analysis combines decision-level evaluation, TF-IDF-based lexical characterization, and embedding-based semantic similarity. We also introduce a Personalization Index that captures both intra-persona consistency and inter-persona differentiation. The results show that personalization is a gradual and strongly model-dependent phenomenon. Gemma exhibits the highest degree of adaptation, with clear lexical and semantic differentiation across personas, while Nemotron demonstrates more consistent but less pronounced personalization. DeepSeek displays weaker and less stable persona adaptation. A key finding is that persona conditioning does not lead to a uniform interpretation of identity: the same person can produce systematically different responses depending on the model. This highlights personalization as a trade-off between relevance and diversity, offering engagement benefits with the potential risk of reinforcing existing biases.