Sumit Kumar Jha

Chain-of-thought (CoT) prompting makes language models write step-by-step explanations, but these steps may not match what the model actually used to choose its answer. Existing faithfulness checks often only test whether changing the written chain changes the answer, without verifying whether th...