Enhancing Confidence Estimates for Black-Box LLMs

The safe deployment of large language models (LLMs) critically depends on accurate uncertainty quantification (UQ). Current UQ methods, such as verbalized confidence or multi-generation approaches, typically operate in a zero-shot manner, generating uncertainty scores without requiring labeled data. However, these methods still necessitate performance evaluation on a specific dataset before deployment. This research proposes an enhancement that capitalizes on this evaluation dataset. By constructing straightforward classifiers, the method predicts the correctness of LLM responses. These classifiers utilize existing zero-shot confidence scores alongside the correctness of similar queries found within the deployment dataset as features. The new approach consistently surpasses the performance of existing UQ scores. Crucially, it achieves this improvement with minimal additional computational cost, making it a practical and efficient upgrade for real-world LLM applications where reliable uncertainty estimates are vital.