My Graph Doesn’t Help! Analyzing Causal Graph Construction, Representation, and Utility in LLMs

Causal reasoning from narrative text requires large language models (LLMs) to recover causal structure and use it effectively. Explicit causal directed acyclic graphs (DAGs) provide a natural representation of such structure, yet a faithful graph may not yield better downstream reasoning. We ask when and how explicit causal structure helps LLM causal reasoning, and to what extent constructing a faithful causal representation and effectively using it are distinct capabilities. On the hard section of CausalProbe-2024, augmenting 3,461 questions with automati cally constructed target-centered causal structure reduces accuracy across all five models. We then build a controlled benchmark around eight four-variable DAG topologies by manually identifying matching structures in recent news articles and representing them as concise narratives, yielding 32 contexts and 160 questions. We compare direct reasoning with ground-truth DAGs, edge-wise counterfactual construction, and holistic one-shot generation, while varying causal-role annota tions and separately measuring reconstruction fidelity and downstream accuracy. Across these analyses, we identify four main findings: explicit causal structure has conditional utility across models and representations, the holistic QuickGraph procedure achieves higher reconstruction $F_1$ than the edge-wise FullGraph pro cedure across all five models, higher reconstruction fidelity is not consistently accompanied by improved downstream reasoning, and causal-role annotations can alter graph utility even when the underlying structure is unchanged. Our results suggest that making pretrained causal knowledge more explicit does not necessarily make it more actionable: representation fidelity and downstream utilization remain distinct capabilities.