Tag
The paper investigates how likelihood ranking and prompting-based evaluation scale differently in LLMs, finding that likelihood accuracy remains stable across model scales while prompted performance improves, suggesting they probe distinct aspects of model behavior.