A research team has built a system that turns a vague data question into a map of every analytical path an AI assistant considered.
The system, called DAG-EDA, lets analysts and a large language model jointly navigate exploratory data analysis instead of just trading questions and answers. It uses two linked structures: an intent graph that breaks an ambiguous question into progressively concrete analysis tasks, letting analysts branch, backtrack, and compare different framings rather than locking into the model's first answer. A second structure, a multi-layered knowledge graph, exposes the LLM's domain assumptions by linking abstract concepts to the actual dataset variables meant to measure them, so a user can check how something like customer satisfaction got mapped to a specific column. Both graphs are built from nothing more than the dataset and the analyst's question, and the resulting analyses render as interactive dashboards. The paper walks through a usage scenario and lays out the design for a planned user study, rather than reporting results from one.
Right now, asking an LLM to explore a dataset usually gets you a single unstructured answer with no record of what else it considered or why one direction won out. That's a real liability for analysts who have to justify conclusions, not just receive them. DAG-EDA's bet is that making the model's reasoning path visible and contestable matters more than making the model smarter.
It's a modest, structural fix rather than a bigger model claim, and it's still a research prototype: the paper proposes a user study but has not yet run one. Whether analysts actually want to audit a graph instead of just asking a follow up question is the open bet here.