Are LLM Decisions Faithful to Verbal Confidence?
This reveals a critical limitation for building trustworthy AI systems, as it shows that calibrated confidence scores do not translate into risk-sensitive behavior, which is an incremental but important finding for AI safety and interpretability.
The study tackled the problem of whether LLMs' expressed confidence aligns with their decision-making under risk, finding that models fail to adjust abstention policies in response to error penalties, leading to utility collapse even when abstention is optimal.
Large Language Models (LLMs) can produce surprisingly sophisticated estimates of their own uncertainty. However, it remains unclear to what extent this expressed confidence is tied to the reasoning, knowledge, or decision making of the model. To test this, we introduce $\textbf{RiskEval}$: a framework designed to evaluate whether models adjust their abstention policies in response to varying error penalties. Our evaluation of several frontier models reveals a critical dissociation: models are neither cost-aware when articulating their verbal confidence, nor strategically responsive when deciding whether to engage or abstain under high-penalty conditions. Even when extreme penalties render frequent abstention the mathematically optimal strategy, models almost never abstain, resulting in utility collapse. This indicates that calibrated verbal confidence scores may not be sufficient to create trustworthy and interpretable AI systems, as current models lack the strategic agency to convert uncertainty signals into optimal and risk-sensitive decisions.