LLMs as Imperfect Experts for Causal Discovery via Argumentation

A recent preprint on arXiv (2602.16481v2) investigates the role of large language models (LLMs) as flawed experts within Causal Assumption-based Argumentation (ABA), a framework for causal discovery through symbolic reasoning. The authors suggest extracting semantic structural priors from variable names and descriptions, which are then combined with evidence of conditional independence. Their experiments on established benchmarks and semantically grounded synthetic graphs yield state-of-the-art results. This research tackles the difficulty of merging data-driven statistical techniques with expert insights, essential for developing sound causal graphs. The study also presents a novel evaluation method for this integration. This work is relevant in AI, machine learning, and causal inference, impacting areas like epidemiology, economics, and social sciences.