An Integrated Architecture for Human-Centric Large Language Models: Multi-Objective Optimization of Personalization, Faithful Explainability, and Linguistic Fairness

Saideh Naderi¹ Reza Gholamrezai²

1) Ph.D. Student, Department of Computer Engineering, Kerman Branch, Islamic Azad University, Kerman, Iran Email:
2) Department of Computer Engineering, Kerman Branch, Islamic Azad University, Kerman, Iran Email:

IOI Code : AAHD-ABCDKF
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    With the expansion of large language models (LLMs) as the backbone of generative AI systems, conventional and non-adaptive approaches face fundamental challenges, including the inability to align with individual user characteristics, lack of transparency in decision-making logic, and performance gaps between high-resource and low-resource languages. This situation underscores the necessity of transitioning from average-centric architectures toward human-centered, adaptable, and responsible models. This study proposes an integrated framework for developing human-centric LLMs by converging three complementary dimensions: multi-level personalization, faithful explainability, and multilingual fairness. Through an analytical–comparative review of recent research, personalization mechanisms at the input, parameter adaptation, and objective alignment levels, explanation generation approaches aligned with the model’s true behavior, and cross-lingual knowledge transfer and few-shot learning techniques for low-resource languages are analyzed and consolidated into a unified architectural framework. The proposed framework formulates the interaction of these three components as a multi-objective optimization problem, where simultaneous optimization of user utility, explanation faithfulness, and reduction of linguistic gaps requires balancing privacy constraints, computational cost, and performance stability. Conceptual analysis indicates that structural convergence of these dimensions can enhance user trust, improve performance in real-world applications, and promote linguistic equity in AI systems.

Keywords : large language models multi level personalization faithful explainability linguistic fairness human centric artificial intelligence