Different Questions, Different Models: Fine-Grained Evaluation of Uncertainty and Calibration in Clinical QA with LLMs

2025-06-12

AI Summary

In high-stakes clinical decision support, large language models (LLMs) require uncertainty estimates that are both accurate and well-calibrated. Method: We systematically evaluate uncertainty modeling capabilities of 10 open-source LLMs across two clinical QA benchmarks, 11 medical specialties, and six question types. We propose a lightweight, single-shot uncertainty estimation algorithm leveraging behavioral signals from reasoning trajectories—bypassing costly sampling while approximating semantic entropy performance—and introduce a multidimensional, fine-grained evaluation framework. Contribution/Results: Our analysis reveals pronounced heterogeneity in calibration performance across specialties and question types—a previously unreported finding. The proposed method reduces expected calibration error by up to 32% across most settings, offering an efficient, reliable, and interpretable uncertainty quantification paradigm for clinical LLM deployment.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Multiagent Systems: Multiagent Systems under Uncertainty

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Accurate and well-calibrated uncertainty estimates are essential for deploying large language models (LLMs) in high-stakes domains such as clinical decision support. We present a fine-grained evaluation of uncertainty estimation methods for clinical multiple-choice question answering, covering ten open-source LLMs (general-purpose, biomedical, and reasoning models) across two datasets, eleven medical specialties, and six question types. We compare standard single-generation and sampling-based methods, and present a case study exploring simple, single-pass estimators based on behavioral signals in reasoning traces. These lightweight methods approach the performance of Semantic Entropy while requiring only one generation. Our results reveal substantial variation across specialties and question types, underscoring the importance of selecting models based on both the nature of the question and model-specific strengths.

Problem

• Evaluating uncertainty estimation methods in clinical QA with LLMs
• Assessing model performance across medical specialties and question types
• Developing lightweight methods for accurate uncertainty estimation

Innovation

• Fine-grained evaluation of uncertainty estimation methods
• Lightweight single-pass estimators using behavioral signals
• Performance comparison across specialties and question types