MetaFaith: Faithful Natural Language Uncertainty Expression in LLMs

TL;DR

This work systematically investigates the faithful calibration problem in LLMs, defined as aligning intrinsic uncertainty with expressed linguistic uncertainty. It introduces a formal, end-to-end evaluation framework based on the F_M metric and cMFG, including a robust LLM-based judge pipeline to quantify decisiveness and intrinsic confidence. The authors show that standard prompts and factual calibration methods fail to achieve faithful uncertainty expression, and that performance varies across models, datasets, and domains. They propose MetaFaith, a metacognition-inspired, prompt-based calibration approach that yields substantial gains (up to 61% in faithfulness) and maintains accuracy, with human evaluations confirming an 83% win rate over a baseline uncertainty elicitation method. The results demonstrate the promise of metacognitive prompting for enhancing trustworthiness and outline directions for broader applicability and future improvements in faithful uncertainty communication.

Abstract

A critical component in the trustworthiness of LLMs is reliable uncertainty communication, yet LLMs often use assertive language when conveying false claims, leading to over-reliance and eroded trust. We present the first systematic study of $\textit{faithful confidence calibration}$ of LLMs, benchmarking models' ability to use linguistic expressions of uncertainty that $\textit{faithfully reflect}$ their intrinsic uncertainty, across a comprehensive array of models, datasets, and prompting strategies. Our results demonstrate that LLMs largely fail at this task, and that existing interventions are insufficient: standard prompt approaches provide only marginal gains, and existing, factuality-based calibration techniques can even harm faithful calibration. To address this critical gap, we introduce MetaFaith, a novel prompt-based calibration approach inspired by human metacognition. We show that MetaFaith robustly improves faithful calibration across diverse models and task domains, enabling up to 61% improvement in faithfulness and achieving an 83% win rate over original generations as judged by humans.

Figures (21)

• Figure 1: Left: Faithful calibration quantifies the alignment between a model's intrinsic uncertainty and expressed uncertainty. Right: Extensive experiments across models and tasks demonstrate that without special instructions (none), LLMs exhibit poor faithful calibration, and generic instructions to express uncertainty (generic) only slightly alleviate this. Our proposed approach (MetaFaith) uses metacognitive prompting to elicit faithful expressions of uncertainty.
• Figure 2: MetaFaith systematically creates metacognitive prompts that can be used to substantially and robustly improve faithful calibration of any instruction-following LLM.
• Figure 3: Plot of linear regression coefficients with 95% confidence intervals for each predictor.
• Figure 4: Efficacy of MetaFaith toward improving faithful calibration of LLMs across models and datasets. Bars report average cMFG across all datasets (values indicated by upper $x$-axis). Average accuracy is denoted by black pointers (values indicated by lower $x$-axis).
• Figure 5: Prompt to extract assertions from model responses.