Guiding model behavior toward specific outcomes by modifying internal representations or training signals.