CArLA

External knowledge

UAI 2026

Leveraging Large Language Models for Causal Discovery: a Constraint-based, Argumentation-driven Approach

Uses language models as imperfect experts that propose causal constraints, then evaluates those constraints alongside statistical evidence through Causal ABA.

The problem

Expert knowledge is valuable for causal discovery but expensive to obtain. Language models can propose semantic relations, yet their outputs may be unreliable or memorised.

The approach

ABAPC-LLM elicits structural constraints from variable names and descriptions, treats the language model as an imperfect expert, and integrates its proposals conservatively with statistical evidence through Causal ABA.

What the work establishes

  • Outperforms both the LLM-only LLM-BFS baseline and the data-only ABA-PC method, reducing average structural Hamming distance by 57% and 17.5%, respectively. This shows that the proposed robust combination recovers more accurate causal graphs.
  • Introduces a new semantically grounded benchmark that randomises graph topology to reduce memorisation bias, with a public graph generator for producing benchmark instances.
  • Achieves the lowest structural Hamming distance and highest recall at every tested graph size, with statistically significant gains over the next-ranked method.

Why it matters to CArLA

This extends CArLA with a scalable source of external knowledge without treating language-model output as unquestionable truth: proposals remain attributable and contestable.

Contributor acknowledgement

  • Zihao Li — author and supervised student contributor

Citation

Li, Z., & Russo, F. (2026). Leveraging Large Language Models for Causal Discovery: a Constraint-based, Argumentation-driven Approach. Proceedings of UAI, PMLR 337, 3631–3667.