GoT-CD Improves Causal Discovery, Reveals Fairness Audit Fragility

A new research paper presents GoT-CD (Graph-of-Thoughts Causal Discovery), a novel approach for inferring directed causal structures from observational data. This method leverages a Graph-of-Thoughts reasoning framework, where the core unit of reasoning is a complete candidate edge set. It generates multiple graphs in parallel, scores them for validity, and merges them under strict constraints to ensure acyclicity, ultimately producing structurally competitive directed acyclic graphs (DAGs) compared to existing large language model (LLM) baselines. Beyond introducing GoT-CD, the study critically examines the reliability of post-hoc path-specific fairness audits. These audits, which assess whether protected attributes influence outcomes through illegitimate pathways, are highly dependent on the accuracy of the underlying causal graph. The research demonstrates that even when a discovered graph exhibits high structural fidelity, it may fail to recover specific pathways crucial for a fairness audit. This can lead to misleading results, where an audit reports no overall effect from a sensitive attribute because the relevant path was missed during discovery, even if mediated effects still exist. The findings underscore the fragility of such audits and the necessity for robust structural discovery alongside fairness analysis.