“Unlike general-purpose LLMs, models like HRM and TRM are task-specific and must be trained individually for each task, such as Sudoku or ARC-Prize.”