METHODOLOGICAL APPROACHES TO PERSONALIZATION IN LARGE LANGUAGE MODELS

Abstract

The article systematizes methodological approaches to the personalization of large language models (LLMs) within the context of their practical application in the field of financial technology. The relevance of this research is driven by the active implementation of LLMs in real-world products, such as government AI assistants, banking chatbots, automated financial advisory systems, and automated agents for processing financial documents. At the same time, base general-purpose LLMs cannot be effectively utilized for niche industry tasks due to a lack of access to up-to-date corporate data, ignorance of specific regulatory frameworks, and an inability to perform actions in external systems. This paper systematizes personalization methods that allow these limitations to be overcome. Five key methods for adapting LLMs to specific business tasks are analyzed: context management and the system prompt – the simplest and most cost-effective way to define the model’s role, style, and behavioral constraints without additional training; prompt engineering, including Chain-of- Thought (CoT) and Tree-of-Thought (ToT) techniques, which significantly improve the accuracy of logical reasoning when solving complex analytical financial tasks; Retrieval-Augmented Generation (RAG), which represents an architectural solution for dynamic access to current data from external sources; Fine-tuning, specifically the parameter-efficient LoRA (Low-Rank Adaptation) method, which allows LLMs to be adapted to highly specialized tasks such as financial document classification, news sentiment analysis, and credit scoring with minimal computational costs; and Function Calling and AI agents, which provide two-way integration of LLMs with external APIs, CRM, and ERP systems. A comparative analysis of these methods was conducted based on five criteria: implementation cost, training data requirements, knowledge currency, integration capabilities with external systems, and the transparency of the generated responses. Based on the analysis conducted, it has been determined that a hybrid architecture combining the use of RAG and Fine-tuning is optimal for integrating AI systems into financial products. Empirical data indicate that each of these methods has a positive impact on response accuracy, and their combination proves to be the most effective. Specific recommendations have been formulated regarding the choice of personalization method depending on the type of application task. Based on the analysis conducted, it has been determined that a hybrid architecture combining the use of RAG and Fine-tuning is optimal for integrating AI systems into financial products. Empirical data indicate that each of these methods has a positive impact on response accuracy, and their combination proves to be the most effective. Specific recommendations have been formulated regarding the choice of personalization method depending on the type of application task.