Scholarly record

COMPARATIVE ANALYSIS OF LLM, LRM, AND DRM MODELS FOR BUILDING AN AI AGENT IN SOCIALLY RELEVANT DOMAINS

Kristina Dineva, Tatiana Atanasova, Ivaylo Keremidarski

Abstract

The growing interest in the development of artificial intelligence (AI) agents is increasingly influencing socially significant domains such as public administration, education, and government services. In these areas, AI systems are expected to achieve high standards of performance, including accuracy, reliability, and contextual awareness, in order to support administrative processes and provide citizens with fast and dependable access to information. For this reason, selecting an appropriate model conception is a critical step in the design of such an agent. Although large language models (LLMs) have demonstrated strong capabilities in natural language understanding and generation, their limitations in reasoning, factual consistency, and integration of real-world data raise important concerns regarding their practical use. At the same time, emerging approaches such as logical reasoning models (LRMs) and deep research models (DRMs) attempt to address some of the weaknesses of conventional language models, while also introducing additional complexity in the design of an AI agent. This paper presents a comparative analysis of the concepts of LLM, LRM, and DRM to identify their respective strengths, limitations, and suitability for different categories of tasks. A set of evaluation criteria is introduced, based on key dimensions such as task complexity, the required level of completeness, the type of response expected by the user, data dependency, and response constraints. The main contribution of this work lies in the analysis of these different model conceptions and the identification of the scenarios in which each of them is most appropriate. Based on this analysis, a decision-making flow is established to guide the selection of a model with the most suitable conceptual foundation. Unlike single-model approaches, the proposed flow supports multi-criteria decision-making and allows for hybrid configurations, aligning model capabilities with the specific requirements of each task. The results indicate that no single model type is universally optimal. Instead, effective AI agent require adaptive strategies that apply the most appropriate conceptual approach to each scenario rather than relying on a one-size-fits-all solution. The findings provide practical guidance for the design of an AI agent in domains that serve diverse groups of users and involve socially significant contexts.

Publication details

Back to publication list