graph-inference

This paper explores integrating large language models as imperfect subject-matter experts into Interpretive Structural Modeling (ISM) to replace labor-intensive expert consensus processes, comparing pairwise, k-wise, rowwise, and full-graph causal discovery methods and finding rowwise (SHD=160, F1=0.77) and full graph (SHD=135, F1=0.73) approaches perform best.

Introduces GraphInfer-Bench, a benchmark to evaluate whether LLMs can perform graph inference—producing open-ended answers about a node and its neighborhood that cannot be retrieved from a single node or path. Experiments show that even frontier LLMs lag behind plain GNNs on these tasks, revealing a capability gap.