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 whether constructing a faithful causal representation and effectively using it are distinct capabilities. On the hard section of CausalProbe-2024, augmenting 3,461 questions with automatically constructed target-centered causal structure reduces accuracy across all five models. We then introduce a controlled benchmark spanning eight four-variable DAG topologies, 32 recent-news 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 annotations and measuring reconstruction fidelity and downstream accuracy. We find that explicit causal structure has conditional utility across models and representations, QuickGraph achieves higher reconstruction $F_1$ than FullGraph 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. This dissociation is clearest for Llama 3.1 8B Instruct, where QuickGraph substantially improves reconstruction fidelity without improving downstream reasoning, highlighting the importance of evaluating structural fidelity and reasoning utility separately in smaller models. These findings suggest that constructing and effectively utilizing causal structure are related but distinct components of LLM causal reasoning, prompting a broader question: what makes a causal representation useful to an LLM beyond simply being structurally correct?