Rescaling Confidence: What Scale Design Reveals About LLM Metacognition

This study investigates the impact of confidence scale design on the metacognitive performance of large language models (LLMs), revealing that conventional 0–100 scales lead to significant discretization of responses. Through systematic manipulation of scale dimensions, the authors find that a reduced 0–20 scale enhances metacognitive efficiency, while boundary compression negatively affects performance. These findings highlight the importance of scale design as a critical factor in evaluating LLM uncertainty, suggesting that practitioners should consider it when developing and assessing models.