CoBe Benchmark for Causal Understanding in Counterfactual Text Editing

The CoBe benchmark evaluates a model's ability to perform counterfactual text editing, requiring it to modify specific causal outcomes while preserving unaffected facts. This distinguishes true causal reasoning from mere statistical association in sensitive domains like medicine.