Tag
This study investigates capability-dependent biases in LLM judges and introduces a calibrated weighted majority voting ensemble method to enhance automated evaluation reliability without requiring labeled data.