l3-unc/CausalDiagnosticity

A Causal Lens for Evaluating Faithfulness Metrics

paper | code

Authors: Kerem Zaman and Shashank Srivastava

This dataset was created to evaluate faithfulness metrics using four controlled tasks: (1) fact-checking, (2) analogy, (3) object counting, and (4) multi-hop reasoning. These tasks assess causal diagnosticity by using counterfactual models with faithful and unfaithful explanations. They are deliberately designed to span varying levels of complexity. The Fact Check task is the simplest, requiring models to answer yes/no questions with minimal reasoning. In contrast, the Analogy task introduces additional complexity through its multiple-choice format and hierarchical, non-mutually exclusive edited relations. The Object Counting task, also a multiple-choice format, goes beyond simple classification by requiring models to demonstrate counting capabilities. Finally, the Multi-hop Reasoning task is the most complex, requiring multiple reasoning steps to arrive at the correct answer, and the most challenging in terms of diagnosticity, as faithful and unfaithful explanation pairs often share significant internal and lexical reasoning components. While the altered models should reason differently, their explanations may not always reference the modifications. To ensure valid faithfulness comparisons, we synthetically generate explanations that emphasize model differences.