Validating the Comparison, Not Just the Scorer: Why Overall Accuracy Can Mislead in Interactive LLM Evaluation
Researchers warn that relying on overall accuracy metrics for LLM evaluators can mask prompt-specific errors.
When evaluating LLM responses to varying conversational pressures, researchers often use automatic scorers validated by a single F1 or accuracy score. This study demonstrates that high aggregate performance can hide significant bias, where a scorer performs worse on specific prompt types, leading to skewed measurements of model behavior.
