Can LLMs Use Linguistic Uncertainty Markers to Reliably Reflect Intrinsic Confidence?
For developers and users of LLMs, this work highlights a fundamental limitation in linguistic confidence expression that undermines trustworthiness and reliability.
LLMs fail to reliably associate epistemic markers (e.g., 'it is likely') with consistent internal confidence levels across distributions, though they preserve some ranking order across tasks. This reveals a persistent miscalibration even under model-centric interpretation.
LLMs' linguistically expressed confidence should faithfully reflect their intrinsic uncertainty. While recent work shows LLMs struggle to use epistemic markers (e.g., "it is likely...") in a human-aligned fashion, it remains unclear whether models can apply their own linguistic confidence framework to associate markers with specific confidence levels in a stable and generalizable way, and how contextual features impact this ability. We conduct the first systematic study of this question, formalizing _marker internal confidence_ (MIC) as the estimated intrinsic confidence a model associates with a specific epistemic marker in a given task domain. We present 7 metrics to evaluate the stability of MICs within and across distributions. Applying our analysis framework to diverse models and tasks, we find that LLMs remain faithfully miscalibrated even under model-centric interpretation of marker meanings, struggling to differentiate markers by internal confidence across distributions despite preserving a somewhat consistent ranking order across tasks. This supplies critical, complementary evidence to existing work toward a holistic understanding of faithful calibration in LLMs, emphasizing the need for more aligned and stable marker use to improve trustworthiness and reliability.