Integrating Physician Diagnostic Logic into Large Language Models: Preference Learning from Process Feedback F i n d i n g s - A C L