PersonaReAct: personalização de respostas em modelos de linguagem com base no Paradigma ReAct

Abstract

Large Language Models (LLMs) have established themselves as powerful tools for natural language generation and information retrieval tasks. However, most of these approaches still lack robust personalization mechanisms, which limits their applicability in contexts where the user’s profile directly influences the usefulness of the response. This work proposes PersonaReAct, a specialization of the ReAct paradigm that incorporates a final adaptation step based on the user’s functional profile. The proposal preserves the original reasoning and action cycle, introducing a personalization function (Tailor) responsible for adjusting the final response according to the user’s characteristics and needs. The model was evaluated through two case studies in distinct domains: the sports domain, using real data from Fortaleza E.C. players, and the medical domain, using hospitalization data from COVID-19 patients. In both contexts, questions associated with different user profiles were considered, and the answers were analyzed based on language, interpretation, and coherence criteria. Several language models with and without explicit reasoning were tested, such as GPT-5 mini and GPT-4.1 mini. The results confirm that the proposed approach generates more appropriate, clear, and contextualized answers when compared to traditional ReAct and ReAct with enriched prompts. It was further concluded that PersonaReAct performs better on models without reasoning and on smaller reasoning-capable models, such as GPT-5 nano. In addition, the modular separation of the personalization step was shown to contribute to greater control and interpretability of the system. The study also discusses limitations and proposes future directions, such as transforming the personalization step into a refinable iterative cycle and exploring practical applications in conversational systems.