MCQA-Eval: Efficient Confidence Evaluation in NLG with Gold-Standard Correctness Labels

TL;DR

MCQA-Eval tackles unreliable correctness labels in NLG confidence evaluation by leveraging gold-standard answers from multiple-choice QA datasets, thereby bypassing the need for a separate correctness function. It reformulates MCQA items into open-form prompts to assess both white-box logit-based and black-box consistency-based confidence measures, enabling scalable, cost-effective evaluation. Across diverse LLMs and QA datasets, MCQA-Eval yields stable rankings that align with traditional methods while removing heavy judgments and biases from correctness labeling. This framework supports calibration-oriented analyses and selective generation, offering practical, scalable benefits for trustworthy NLG in high-stakes domains.

Abstract

Large Language Models (LLMs) require robust confidence estimation, particularly in critical domains like healthcare and law where unreliable outputs can lead to significant consequences. Despite much recent work in confidence estimation, current evaluation frameworks rely on correctness functions -- various heuristics that are often noisy, expensive, and possibly introduce systematic biases. These methodological weaknesses tend to distort evaluation metrics and thus the comparative ranking of confidence measures. We introduce MCQA-Eval, an evaluation framework for assessing confidence measures in Natural Language Generation (NLG) that eliminates dependence on an explicit correctness function by leveraging gold-standard correctness labels from multiple-choice datasets. MCQA-Eval enables systematic comparison of both internal state-based white-box (e.g. logit-based) and consistency-based black-box confidence measures, providing a unified evaluation methodology across different approaches. Through extensive experiments on multiple LLMs and widely used QA datasets, we report that MCQA-Eval provides efficient and more reliable assessments of confidence estimation methods than existing approaches.

Figures (7)

• Figure 1: Illustration of the existing evaluation framework (blue) vs our proposed MCQA-Eval (green). Unlike existing frameworks, we avoid the costly and unreliable correctness function module by using multiple-choice datasets. This requires slight modification to the confidence estimation steps, which is elaborated in \ref{['sec:method']}.
• Figure 2: The AUROC ranking of black-box confidence measures (on LLaMA2-13B and CoQA) is sensitive to the threshold $\tau$.
• Figure 3: Using LLM judges as $f$, while flexible, still has inherent noise. Different LLMs disagree on whether a response is correct (left). Even the same LLM (GPT, right) could deliver different opinions simply due to the randomness in generation.
• Figure 4: We reformat each option from the multiple-choice question (left), by injecting the option to a free-form QA prompt. One could typically apply any confidence estimation method by treating this option as if it was generated by the base LM. For black-box confidence measures that require additional responses, we only feed the prompt to the base LM.
• Figure 5: (a) and (b) show the performance of 4 different LLMs and 12 different confidence estimation methods on the C-QA dataset. A higher AUROC indicates better performance.