Research

Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps

Proposes FUR, which measures chain-of-thought faithfulness by unlearning reasoning steps from model parameters and checking whether the model's prediction changes, across four LMs and five MCQA datasets.