Measuring CoT Faithfulness by Unlearning Reasoning Steps

"Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps" (EMNLP 2025 Outstanding Paper): introduces the FUR (Faithfulness by Unlearning Reasoning steps) metric, which erases individual reasoning steps from a model's parameters to test whether a chain of thought is parametrically faithful — i.e., whether the model actually relies on the steps it verbalizes. Technion-led by Yonatan Belinkov.