Proposta de Tese de Doutorado de Bruno Gomes Tavares dos Santos

Proposta de Tese de Doutorado

Aluno(a): Bruno Gomes Tavares dos Santos
Orientador(a): Joao Luiz Dihl Comba

Título: Framework for scientific hypothesis generation by leveraging LLMs
Linha de Pesquisa: Mineração, Integração e Análise de Dados

Data: 24/08/2026
Hora: 08:30
Local: Esta banca ocorrerá de forma híbrida (virtual e presencial), na sala 216/43412 do Instituto de Informática/UFRGS e pelo link https://us02web.zoom.us/j/86798203933?pwd=NIzLUdTd4y3Vf5Gc5hbPgPPn8pjLUi.1.

Banca Examinadora:
-Viviane Pereira Moreira (UFRGS)
-Márcia Helena Barbian (UFRGS)
-Emanuele Marques Rodrigues Santos (UFC)

Presidente da Banca: Joao Luiz Dihl Comba

Resumo: Exploratory Data Analysis (EDA) has a fundamental role in scientific discovery and data-driven decision making. However, the growing volume and complexity of modern datasets make it increasingly difficult for data analysts to manually explore the search space and identify meaningful patterns. While Large Language Models (LLMs) have recently demonstrated remarkable capabilities in reasoning and knowledge synthesis, their direct application to hypothesis generation remains challenging due to hallucinations, limited statistical grounding, and the absence of systematic exploration strategies. In this sense, this research proposes a multi-agent framework, designed to support hypothesis-driven exploratory data analysis. Furthermore, we propose HELIOS (Hypothesis Exploration with LLM-driven Intelligent Orchestration System), a Visual Analytics tool that accommodates the framework. The proposed approach combines LLM-based reasoning with formal statistical validation, allowing autonomous agents to generate hypotheses, evaluate them through appropriate statistical tests, synthesize interpretable narratives, and recommend subsequent analytical directions. Rather than replacing human analysts, the system is designed to function as an analytical partner, accelerating exploration while maintaining statistical rigor and interpretability. Preliminary experiments demonstrate that carefully engineered prompts substantially improve hypothesis quality. Additionally, our results also indicate that narrative generation and next-steps agents contribute to the interpretability of the analytical process by translating statistical findings into actionable insights. The expected contributions of this research include a statistically grounded and open-source multi-agent architecture for hypothesis generation, as well as a study about integrating LLMs into exploratory data analysis workflows. Finally, this work also establishes the foundation for future extensions involving adaptive context engineering, knowledge graph integration, and exploration–exploitation strategies for analytical systems.

Palavras-Chave: Exploratory Data Analysis. LLMs. Multiagent. Hypothesis generation. Data analysis.