A Context-Aware Dual-Metric Framework for Confidence Estimation in Large Language Models

Accurate confidence estimation is essential for trustworthy large language models (LLMs) systems, as it empowers the user to determine when to trust outputs and enables reliable deployment in safety-critical applications. Current confidence estimation methods for LLMs neglect the relevance between responses and contextual information, a crucial factor in output quality evaluation, particularly in scenarios where background knowledge is provided. To bridge this gap, we propose CRUX (Context-aware entropy Reduction and Unified consistency eXamination), the first framework that integrates context faithfulness and consistency for confidence estimation via two novel metrics. First, contextual entropy reduction represents data uncertainty with the information gain through contrastive sampling with and without context. Second, unified consistency examination captures potential model uncertainty through the global consistency of the generated answers with and without context. Experiments across three benchmark datasets (CoQA, SQuAD, QuAC) and two domain-specific datasets (BioASQ, EduQG) demonstrate CRUX’s effectiveness, achieving the highest AUROC than existing baselines.

💡 Research Summary

The paper addresses a critical gap in confidence estimation for large language models (LLMs) used in contextual question‑answering (CQA) tasks. Existing methods focus on either self‑consistency across multiple generations or self‑assessment via prompting, but they ignore whether the generated answer actually aligns with the supplied context. This oversight can lead to “consistent but ungrounded” answers, which is especially problematic in safety‑critical applications where users must know when to trust a model’s output.

To remedy this, the authors propose CRUX (Context‑aware entropy Reduction and Unified consistency eXamination), the first framework that jointly evaluates two complementary dimensions: (1) contextual faithfulness and (2) model uncertainty. The first dimension, Contextual Entropy Reduction (ΔH), quantifies how much the presence of context reduces the entropy of the answer distribution. The procedure samples n answers conditioned on (question + context) and n answers conditioned on the question alone. Answers are clustered using bidirectional entailment, producing semantic clusters K(c,q) and K(q). Entropy is computed for each set of clusters, and ΔH = H(K(q)) − H(K(c,q)). A large positive ΔH indicates that the context provides substantial information, shrinking the hypothesis space; a near‑zero ΔH suggests either that the model already knows the answer (low model uncertainty) or that the model fails to use the context (high model uncertainty).

The second dimension, Unified Consistency (GC), disambiguates the ΔH≈0 case. All 2n answers (contextual and context‑free) are placed in a graph where nodes are answers and edge weights are semantic similarity scores. Two possible GC formulations are offered: average pairwise distance or distance from the cluster centroid. High GC (low dispersion) means the answer set is stable regardless of context, implying low model uncertainty. Low GC (high dispersion) signals that the model’s internal knowledge is insufficient or that it cannot effectively incorporate the context.

CRUX fuses ΔH and GC via a lightweight two‑layer multilayer perceptron (MLP). The concatenated vector