JMIR Publications

ABSTRACT

This research introduces LIBR-TU Research Agent (LITURAt), a novel AI-powered system designed to facilitate the exploration of relationships within scientific datasets, leveraging Large Language Models (LLMs) and an agent-based architecture. By integrating local dataset analysis with relevant literature retrieval, LITURAt enables users of all expertise levels to extract insights from complex datasets. Our system employs a plan-and-solve approach, combining dynamic calculations and literature contextualization to answer user queries with high accuracy. Experiments using the Tulsa 500 dataset demonstrated strong internal and external consistency, with over 90% agreement in results, and an accuracy rating of 80.3% in alignment with GPT-4 evaluations. LITURAt’s ability to break down complex queries into manageable steps allows for the analysis of multi-variable relationships and predictive capacities, making it versatile in addressing a wide range of scientific questions. However, limitations such as model stability during highly complex tasks, dependence on dataset quality, and the inherent challenges of pre-trained LLMs must be addressed. Despite these challenges, LITURAt represents a significant advancement in democratizing access to scientific data, offering a promising tool for researchers and non-experts alike to engage with and derive insights from experimental data. Future enhancements, including improved summarization techniques and the incorporation of multiple data sources, will further extend its utility across diverse scientific disciplines.