Causal-Audit: Auditable Causal Reasoning for LLMs

Advancing large language models (LLMs) beyond surface-level correlations to understand underlying causal mechanisms is crucial for true reasoning. However, current LLM-based methods for causal and intervention-based question answering often rely on implicit language-level reasoning, leading to opaque assumptions, unverifiable paths, and fragile predictions, especially without context. To address these limitations, researchers propose Causal-Audit, an explicit and auditable causal reasoning framework. This method reframes causal inference as structured reasoning over an explicit causal graph, breaking it down into four modular stages instead of an end-to-end prediction. A key innovation is its target-aware causal graph construction, which uses the target variable as a constraint to suppress irrelevant variables and noise during graph expansion. Causal-Audit also introduces a path-level causal evidence aggregation mechanism. This combines multiple causal paths while accounting for both reinforcing and counteracting effects, leading to more robust decision-making. Extensive experiments on three benchmarks demonstrate that this framework consistently outperforms existing LLM-based methods, providing interpretable and auditable causal reasoning traces.