LLM-Guided Causal Discovery Reveals Fairness Pathways

A new preprint (v2, Jan 7, 2026) by Khadija Zanna proposes a hybrid LLM-guided causal discovery framework that extends breadth-first search with active learning and dynamic composite scoring. The method prioritizes variable-pair queries using mutual information, partial correlation, and LLM confidence and is evaluated on a semi-synthetic UCI Adult benchmark embedding bias pathways and confounders. Results show LLM-guided variants outperform baselines at recovering fairness-relevant structures under noisy conditions.