Role-Prompting Fails to Align Model Capabilities with Assigned Personas

The RoleCapBench framework demonstrates that reasoning models maintain high-level capabilities even when prompted to assume low-capability personas, such as a kindergartner. Evaluations show that while stylistic voice alignment is successful, underlying mathematical and reasoning proficiency remains decoupled from the assigned role.