Researchers propose Causal-Audit, a framework for explicit and auditable graph-based causal reasoning in LLMs

Existing LLM-based causal and intervention-based question answering methods often rely on implicit, language-level reasoning, which leads to opaque causal assumptions, unverifiable reasoning paths, and fragile predictions—especially in context-free settings where prior context is absent.

The proposed Causal-Audit framework reformulates causal inference as structured reasoning over an explicit causal graph, implemented through four modular stages. This approach contrasts with end-to-end prediction, making both the reasoning process and assumptions more transparent and auditable.

A central innovation is a target-aware causal graph construction strategy that treats the target variable as a core constraint during graph expansion. This design helps suppress irrelevant variables, spurious causal relations, and reasoning noise, focusing the model’s inference on the most pertinent causal pathways.

The framework also introduces a path-level causal evidence aggregation mechanism that combines multiple causal paths while modeling both reinforcing and counteracting effects. This enables more robust decision-making than single-chain reasoning, which can be brittle or misleading in complex causal scenarios.

In experiments across three benchmarks, the Causal-Audit framework consistently outperformed existing LLM-based methods. The authors report that the approach not only improves performance but also provides interpretable and auditable causal reasoning traces, addressing a longstanding challenge in explainable AI.