Epistemic Verbalization in Uncertainty Communication

Updated 4 May 2026

Epistemic verbalization is the explicit linguistic mapping of uncertainty and belief using modals and hedges to signal graded confidence.
It integrates formal semantics, Bayesian models, and computational metrics to quantify and operationalize uncertainty in language.
This framework enhances understanding in game-theoretic semantics, decision theory, and AI reasoning through empirical and computational studies.

Epistemic verbalization is the explicit linguistic externalization of uncertainty, belief strength, or evidence-basis regarding the truth of a proposition, typically realized through epistemic modals, hedge tokens, graded attitude verbs, and meta-cognitive phrases. This phenomenon spans foundational logic, decision theory, language typology, cognitive science, and the training or evaluation of LLMs. Across these domains, epistemic verbalization serves both as an informational signal in communication and as a computational mechanism for managing and expressing uncertainty, shaping coordination, reasoning, and the evaluation of agent belief states.

1. Formal and Descriptive Definitions

Epistemic verbalization, as established in game-theoretic semantics and philosophy of language, is the overt mapping of partial or graded confidence states into specific linguistic forms—most canonically epistemic modals such as “might,” “possibly,” or “probably.” These modals articulate the speaker's doxastic degree of belief in $p$ , extending the common ground not only with propositional content but with meta-information on the associated credence or evidential status (Sbardolini, 2021, Li et al., 2 Jun 2025). In the context of LLMs and reasoning models, epistemic verbalization operationally includes tokens and utterances directly expressing uncertainty or alternative hypotheses, such as “wait,” “hmm,” “perhaps,” “might,” “likely,” and related lexical markers (Kim et al., 25 Mar 2026, Kim et al., 16 Mar 2026).

A typological framework positions these expressions along two axes: evidential basis (direct, inferred, hearsay) and commitment strength (from complete uncertainty to full certainty), with realized forms including modal auxiliaries (“may,” “might,” “must”), attitude verbs (“doubt,” “believe,” “know”), and graded probability words (“possible,” “likely,” “certain”) (Li et al., 2 Jun 2025).

2. Formal Semantics: Relational, Bayesian, and Hybrid Accounts

Epistemic modalities are modeled in Kripke-style relational semantics by introducing a set of possible worlds $W$ , an accessibility relation $R \subseteq W \times W$ representing epistemic compatibility, and a valuation $V$ for atomic propositions. The modal “Might $p$ ” (written $\Diamond p$ ) holds at $w$ iff there is some $w' \in W$ with $w R w'$ and $M, w' \models p$ —that is, $W$ 0 is epistemically accessible or possible given the speaker’s state (Sbardolini, 2021). In dynamic epistemic logic, modal expressions and announcements update Kripke models, narrowing the common ground and changing the epistemic landscape for all agents involved (Sileo et al., 2023).

A Bayesian formulation assigns each agent a probability function $W$ 1 over $W$ 2 and defines thresholds for outright assertion or hedged assertion. For example, asserting “Might $W$ 3” corresponds to $W$ 4 lying within $W$ 5 for scenario-dependent values, with outright assertion reserved for higher thresholds (Sbardolini, 2021, Li et al., 2 Jun 2025). This quantifies linguistic choices directly in terms of subjective degree of belief and rational action under uncertainty.

3. Measurement and Operationalization in Computational Systems

In natural language processing and computational modeling, epistemic verbalization is empirically measured by counting the frequency of selected uncertainty tokens within generated multi-step reasoning chains, known as chain-of-thought (CoT) outputs. Metrics such as $W$ 6 (with $W$ 7 a fixed set of epistemic markers) provide a quantitative basis, augmented by conditional mutual information $W$ 8 to capture the reduction of uncertainty provided by auxiliary context $W$ 9 (Kim et al., 25 Mar 2026, Kim et al., 16 Mar 2026).

In political discourse or large text corpora, the Evidence–Minus–Intuition (EMI) score combines LLM ratings (on a 0-4 scale for evidence- or intuition-based reasoning) with embedding-based semantic similarity to domain-specific anchors. The resulting hybrid EMI provides a scalable, robust index of how much epistemic verbalization pervades massive text corpora, facilitating diachronic and cross-national analyses in domains such as parliamentary speech (Aroyehun et al., 21 Apr 2026).

4. Cognitive and Theory-of-Mind Modeling

Epistemic verbalization is central not only in external communication but for modeling how agents reason about beliefs and knowledge, including false belief, graded belief, and mutual ignorance. In Bayesian theory-of-mind (ToM) frameworks, natural language is translated into a formal “epistemic language of thought” (ELoT), where modal qualifiers are directly mapped to probability thresholds (e.g., “might $R \subseteq W \times W$ 0” iff $R \subseteq W \times W$ 1). Empirical findings show that such models predict human plausibility judgments for modal claims and belief attributions much better than unconstrained neural models (Ying et al., 2024). In dynamic epistemic logic (DEL), controlled mappings between formal structure and verbalization make the faithful generation and interpretation of epistemic language tractable for both human and AI performance benchmarking (Sileo et al., 2023).

5. Role in Reasoning and Decision-Making Systems

Within LLMs, epistemic verbalization acts as an explicit control mechanism during stepwise reasoning, distinguishing between procedural continuation (executing sub-tasks) and points of epistemic assessment. Information-theoretically, externalizing uncertainty (formally, increasing $R \subseteq W \times W$ 2, where $R \subseteq W \times W$ 3 is latent uncertainty) directly supports productive information gain when procedural reasoning has reached local stagnation (Kim et al., 16 Mar 2026). The empirical suppression of epistemic verbalization—often a byproduct of self-distillation or exposure to overly rich context—leads to collapsed out-of-distribution performance, with up to 40% loss in benchmarked OOD settings (Kim et al., 25 Mar 2026). Distillation, combined with aggressive pruning of uncertainty markers, removes the very signals required for downstream self-correction and adaptive hypothesis revision.

6. Applications and Empirical Domains

Epistemic verbalization structures coordination under information asymmetry. Game-theoretic models, such as the two-player coordination game (beach vs. café scenario) (Sbardolini, 2021), show that strategic hedging with “might $R \subseteq W \times W$ 4” induces iterative updating of mutual expectations and increases the equilibrium probability of successful joint action compared to outright assertion or silence. In group settings and reflective decision-making, structured protocols—combining visualization of latent preference clusters, explicit “switches” for weak preferences, and guided discussion—surface unconscious epistemic states, facilitating both metacognitive awareness and collective articulation of latent beliefs (0803.4074).

In societal applications, the degree and pattern of epistemic verbalization in parliamentary speech is predictive of deliberative democratic quality and legislative transparency, as shown by longitudinal EMI measures across millions of parliamentary records (Aroyehun et al., 21 Apr 2026).

7. Limitations, Recommendations, and Future Directions

Substantial limitations persist in both the generation and recognition of epistemic verbalization. Empirical studies show LLMs frequently miscalibrate modal expressions, over-committing when they should hedge, or failing to employ skepticism or doubt even when warranted by contradictory evidence (Li et al., 2 Jun 2025). Robust epistemic verbalization requires explicit linkage between evidence, subjective probability, commitment thresholds, and linguistic realization.

Recommendations include: (a) explicit fine-tuning or multi-task learning to teach models graded hedges, evidential distinctions, and surface-syntax realization; (b) regularization strategies to preserve or reward epistemic uncertainty signals during distillation or reasoning (Kim et al., 25 Mar 2026); (c) the use of hybrid measurement scores for corpus-level benchmarking (Aroyehun et al., 21 Apr 2026); and (d) controlled frameworks for translating formal epistemic content to high-fidelity verbalization in both generation and comprehension (Sileo et al., 2023, Ying et al., 2024).

Future research directions emphasize integrating multi-modal evidence (vision+text), developing more granular calibration and uncertainty metrics, and embedding epistemic objectives directly in model learning loops to achieve tractable, transparent, and robust epistemic verbalization in both artificial and human discourse.

Key Paper References:

(Sbardolini, 2021) Epistemic Modality and Coordination under Uncertainty
(0803.4074) Reflective visualization and verbalization of unconscious preference
(Kim et al., 25 Mar 2026) Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?
(Aroyehun et al., 21 Apr 2026) Epistemic orientation in parliamentary discourse is associated with deliberative democracy
(Li et al., 2 Jun 2025) Representations of Fact, Fiction and Forecast in LLMs: Epistemics and Attitudes
(Kim et al., 16 Mar 2026) Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty
(Sileo et al., 2023) MindGames: Targeting Theory of Mind in LLMs with Dynamic Epistemic Modal Logic
(Ying et al., 2024) Understanding Epistemic Language with a Language-augmented Bayesian Theory of Mind