Externalized CPDAG Summaries Improve LLM Causal Deduction
Wentao Sun, João Paulo Nogueira, Dominique Verchere, Mathieu Acher, Alonso Silva
原文摘要
Observational data rarely identify a single causal structure. Such structures are typically modeled as directed acyclic graphs (DAGs) whose arrows point from direct causes to their effects, and observed correlations and conditional independencies usually leave a whole family of DAGs that fit the data equally well. A causal claim follows from the data only if it holds in every one of them. Classical causal discovery methods such as the Peter--Clark (PC) algorithm summarize this family as a completed partially directed acyclic graph (CPDAG): an edge is directed when every compatible graph orients it the same way, and undirected when the data leave its direction open. On the Corr2Cause benchmark, large language models (LLMs) that reason in free text often lose track of this family and decide from a few local edges. We propose \emph{Structured Thinking}, a two-turn method in which the model first summarizes the CPDAG in a constrained schema and then checks the claim against that summary. Without fine-tuning or an external solver, it raises Qwen3.5-27B from $77.5$ to $86.5$ F$_1$ (mean of three seeds) over a baseline that receives explicit PC-algorithm instructions, whereas a second turn of free-text analysis does worse than that baseline. The gain holds on test splits with renamed variables or paraphrased hypotheses and on other models. Given only the graph and the claim, the model's answer flips when the graph is replaced by a valid CPDAG under which the claim has the opposite truth value. Moreover, replacing the model's graphs with reference CPDAGs on a fixed subset shows that most remaining errors arise in building the graph rather than in reading it. Making causal structure explicit can help LLMs reason more reliably, provided the intermediate representation is itself trustworthy.
cs.AI