Can large language models perform causal discovery?

Large language models can suggest which variables plausibly influence each other by reading variable names and descriptions, but asking one to generate an entire graph is hard to constrain and tends to break down on context, validity, or cost as graphs grow. Research on Bayesian network structure learning, such as ABSOL, suggests LLMs work better as bounded advisers inside a statistical pipeline, for example ranking candidate parents or choosing which variables to sample together, with every suggestion checked against the data and the graph’s structural rules. Reading the learned edges as causes still requires additional assumptions beyond the data.