Banca de DEFESA: REJANIO DAS NEVES DE MORAES FILHO

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : REJANIO DAS NEVES DE MORAES FILHO
DATE: 28/08/2026
TIME: 09:00
LOCAL: Remoto
TITLE:

State Graph-Based Architecture for Automotive Recommender Systems using Large Language Model

KEY WORDS:

Recommendation Systems; Large Language Models; LangGraph; ReAct; Generative Artificial Intelligence; Retrieval-Augmented Generation; Automotive Recommendation.

PAGES: 100
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

Large Language Models (LLMs) have significantly expanded the capabilities of recommendation systems by enabling the interpretation of natural language queries and the generation of context-aware responses. However, applications in specialized domains, such as automotive recommendation, still face challenges related to the high computational cost of deep reasoning, the instability of orchestration architectures based on autonomous agents, and the degradation of response quality caused by small variations in query formulation. In this context, this work proposes and evaluates an automotive recommendation architecture based on State Graphs (LangGraph), comparing its performance with the traditional ReAct orchestration paradigm. The proposed architecture integra-tes information from user profiles, vehicle technical specifications, and accident history databases, enabling recommendations grounded in multiple contextual sources. The experimental evaluation was conducted using seven locally deployed language models ranging from 4 to 72 billion parameters, considering five queries with different levels of complexity, as well as sensitivity and ablation experiments. The architectures were eva-
luated using computational metrics, including latency, token consumption, throughput, verbosity, and reasoning density, in addition to qualitative metrics based on accuracy, BERTScore, and a unified final score. The results demonstrate that the State Graph-based architecture substantially reduces execution failures, improves recommendation stability, and decreases the sensitivity of language models to both the ordering of query constraints and variations in users’ demographic information. Furthermore, smaller language models orchestrated by the proposed architecture achieved qualitative performance comparable to or better than that of significantly larger models operating under the ReAct paradigm. The ablation experiments also showed that compartmentalizing the execution flow and reducing the cognitive load associated with peripheral information contribute to improving recommendation quality in more complex queries. The findings indicate that State Graph-based orchestration provides an effective approach for improving the robustness, quality, and computational efficiency of LLM-based recommendation systems, demonstrating that execution flow engineering can have an impact as significant as the scale of the language models themselves.

COMMITTEE MEMBERS:
Presidente - 2885532 - IVANOVITCH MEDEIROS DANTAS DA SILVA
Externo ao Programa - 2249146 - CARLOS MANUEL DIAS VIEGAS - UFRNExterno à Instituição - JUAN MOISES MAURICIO VILLANUEVA - UFPB
Externa à Instituição - MARIANNE BATISTA DINIZ DA SILVA - UFAL