RCD

A training-free decoding method that contrasts token distributions obtained with and without a reasoning process, thereby strengthening the model's reliance on reasoning content and mitigating the query shortcut problem in chain-of-thought prompting.