The method asks the model to choose a task-specific role before producing the final answer.

The project studies how self-selected roles affect zero-shot reasoning performance across arithmetic, commonsense, and strategy QA benchmarks.