A Causal Lens for Evaluating Faithfulness Metrics
Keywords: faithfulness, diagonsticity, natural language explanations, interpretability, model editing
Abstract: The increasing capabilities of Large Language Models (LLMs) have made natural language explanations a promising alternative to traditional feature attribution methods for model interpretability. However, while these explanations may seem plausible, they can fail to reflect the model's underlying reasoning faithfully. The idea of faithfulness is critical for assessing the alignment between the explanation and the model's true decision-making mechanisms. Although several faithfulness metrics have been proposed, they lack a unified evaluation framework. To address this limitation, we introduce Causal Diagnosticity, a new evaluation framework for comparing faithfulness metrics in natural language explanations. Our framework extends the idea of diagnosticity to the faithfulness metrics for natural language explanations by using model editing to generate faithful and unfaithful explanation pairs. We introduce a benchmark consisting of three tasks: fact-checking, analogy, and object counting, and evaluate a diverse set of faithfulness metrics, including post-hoc explanation-based and chain-of-thought (CoT)-based methods. Our results show that while CC-SHAP significantly outperforms other metrics, there is substantial room for improvement. This work lays the foundation for future research in developing more faithful natural language explanations, highlighting the need for improved metrics and more reliable interpretability methods in LLMs.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9278
Loading