Performance of causal foundation models across identification regimes

Determine how the performance of causal foundation models changes when the information available for causal identification changes across observational views.

Background

Causal foundation models are designed to estimate causal effects from observational datasets under varying structural causal models and identification assumptions. Different observational views expose estimators to different identifying information, such as observed confounders, mediators, instruments, or confounding proxies. The paper notes that existing evaluations use differing pre-training environments and protocols, making it difficult to isolate the effect of changes in available identifying information. The unresolved issue is whether and how causal foundation model performance changes when the underlying causal world and target effect are held fixed while only the observational view changes.