Listening to the Wise Few: Query-Key Alignment Unlocks Latent Correct Answers in Large Language Models
The research introduces novel scoring methods, QK-score and Attention Score, derived from specific 'select-and-copy' attention heads in LLMs. These scores focus on internal model representations rather than standard formats to find correct answers in multiple-choice questions. This approach shows significant performance gains across various benchmarks by leveraging stable, universal model components.