Causal-Audit makes LLM reasoning transparent with explicit causal graphs

Large language models struggle with causal reasoning, often producing opaque and unverifiable predictions when faced with complex intervention-based questions. Enter Causal-Audit, a new framework from researchers at Griffith University (Su Lan, Xuefei Yin, Yanming Zhu, Alan Wee-Chung Liew) that makes the reasoning process explicit and auditable. The system breaks causal inference into four modular stages, constructing a structured causal graph instead of relying on implicit end-to-end prediction. A key innovation is its target-aware causal graph construction, which uses the target variable as a core constraint during graph expansion to eliminate irrelevant variables and spurious correlations.

Causal-Audit also introduces a path-level causal evidence aggregation mechanism that combines multiple causal paths, modeling both reinforcing and counteracting effects to enable robust decision-making beyond single-chain reasoning. Extensive experiments on three benchmark datasets show the framework consistently outperforms existing LLM-based methods while providing interpretable, auditable reasoning traces. The paper has been accepted for the 64th Annual Meeting of the Association for Computational Linguistics (ACL) in 2026, signaling strong peer recognition of its approach.