SGHA Discovers Research Problems Using Local LLMs

The early stages of scientific research, particularly problem formulation, often rely on proprietary large language models (LLMs), which pose challenges regarding transparency, auditability, and data privacy. These models can also introduce biases or hallucinations due to their opaque parametric knowledge. To address these issues, the Structural Gap Hypothesis Agent (SGHA) has been developed as a fully automated, corpus-first system for discovering research problems. SGHA operates entirely on a local LLM, structuring scientific literature into evidence-linked paper objects and a typed evidence graph. This system identifies unresolved structural patterns across papers, screens potential gaps, and then formulates traceable research problem families, complete with assumptions, objectives, and success criteria. By running on a local, open-weight 9B language model, SGHA offers an auditable and secure alternative to frontier models for scientific discovery, demonstrating promising results in various machine learning domains.