Faithful Calibration (FC) in Large Language Models

• Faithful Calibration (FC) is the practice of aligning a model’s expressed natural language uncertainty with its intrinsic confidence, independent of factual accuracy.
• FC methodologies use metrics that compare linguistic decisiveness with internal confidence estimates, often employing sampling-based evaluations and reinforcement learning.
• FC enhancements improve model reliability in reasoning by ensuring metacognitive self-assessment, though challenges remain in step segmentation and estimator disagreement.

Faithful Calibration (FC) is the alignment between a model’s intrinsic confidence and its expressed confidence, especially when confidence is communicated in natural language rather than as a scalar probability. In the recent LLM literature, FC is treated as a distinct reliability target: a response is faithful when linguistic decisiveness or self-reported confidence matches the model’s own internal uncertainty, not merely empirical correctness frequencies. This framing appears explicitly in work on faithful confidence calibration for natural-language uncertainty expression, metacognitive reinforcement learning, and large reasoning models, where FC is analyzed at assertion, sentence, and reasoning-step granularity and is separated from standard factual calibration, accuracy, and answer correctness (Liu et al., 30 May 2025, Liu et al., 30 Jun 2026, Gani et al., 2 Jun 2026).

1. Definition and conceptual boundaries

In direct FC work, the central object is not whether a model is correct, but whether it communicates uncertainty in a way that faithfully reflects its own uncertainty state. MetaFaith defines the target as making the natural-language expression of uncertainty faithfully reflect the model’s own intrinsic uncertainty, and emphasizes that faithfulness and factuality are distinct axes. A response can therefore be false yet faithfully hedged if the model was genuinely uncertain, and correct yet unfaithfully hedged if the model sounded uncertain despite high internal confidence (Liu et al., 30 May 2025).

This distinguishes FC from standard calibration. In the FC framing used by metacognitive RL, factual calibration aligns confidence with empirical accuracy, whereas FC aligns expressed confidence with intrinsic confidence. The same paper states the goal as enabling models to express uncertainty that genuinely reflects their estimated intrinsic confidence. It therefore treats FC as fundamentally metacognitive: the model must monitor its own performance and then express that self-assessment faithfully (Liu et al., 30 Jun 2026).

The distinction becomes sharper in large reasoning models (LRMs). There, FC is defined as the alignment between intrinsic confidence and linguistically expressed confidence throughout a chain-of-thought trace. This is explicitly not reducible to answer accuracy, standard calibration, or final-answer verbal confidence. The motivating concern is that long reasoning traces are often interpreted by users as evidence of deliberation, competence, and confidence, even when the wording does not track the model’s own internal uncertainty (Gani et al., 2 Jun 2026).

2. Formalizations and evaluation metrics

Existing FC formalisms share a common template: faithfulness is high when the distance between an internal confidence proxy and an external confidence signal is small. In MetaFaith, a response is decomposed into assertions $R=\{A_1,\ldots,A_N\}$, and the example-level score is

$$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$

where $dec(A_n)$ is linguistic decisiveness and $conf_M(A_n)$ is intrinsic confidence. Misalignment in either direction counts as unfaithfulness: overconfident language when intrinsic confidence is low, or underconfident language when intrinsic confidence is high (Liu et al., 30 May 2025).

RLMF adopts the same structure at sentence level. For a response $R=\{s_1,\ldots,s_L\}$, it defines

$$F^M_{Q, R} := 1 - \frac{1}{L}\sum_{l=1}^{L} |conf_M^{expressed}(s_l) - conf_M^{intrinsic}(s_l) |.$$

Here the expressed side may be numerical or linguistic, depending on stage: Stage 1 trains numerical self-reported confidence, while Stage 2 maps those calibrated scores into natural-language hedging (Liu et al., 30 Jun 2026).

The LRM framework generalizes the same idea to reasoning traces. For a trace $T=(s_1,\ldots,s_n)$, step-level faithfulness is

$F(s_i) = 1 - \lvert D(s_i) - C(s_i) \rvert,$

and trace-level faithfulness is

$$F_C(T) = 1 - \frac{1}{|\mathcal{I}(T)|} \sum_{i \in \mathcal{I}(T)} \lvert D(s_i) - C(s_i) \rvert.$$

The same paper also defines trace-level mean confidence

$$C(T) = \frac{1}{|\mathcal{I}(T)|} \sum_{i \in \mathcal{I}(T)} C(s_i),$$

which becomes the conditioning variable for dataset-level summaries (Gani et al., 2 Jun 2026).

At dataset level, early FC work uses conditional Mean Faithfulness Generation ($$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$0). RLMF reproduces the standard definition

$$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$1

together with an equal-width-bin implementation. Both RLMF and the LRM paper then argue that this metric is biased when intrinsic confidence occupies only part of $$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$2, and introduce the width-weighted, equal-mass-bin variant

$$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$3

where $$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$4 is the intrinsic-confidence-axis width of bin $$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$5. In both papers, the motivation is to evaluate FC uniformly over the model’s empirical confidence support rather than penalizing it for not occupying unused regions of $$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$6 (Liu et al., 30 Jun 2026, Gani et al., 2 Jun 2026).

3. Estimating intrinsic confidence and expressed confidence

A recurring difficulty in FC is that intrinsic confidence is not observed directly. MetaFaith and RLMF therefore operationalize it through sampled-response consistency. In MetaFaith, for each assertion $$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$7, the model samples $$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$8 additional responses and checks whether each sampled response supports the assertion. The inconsistency scores are mapped as

$$F_M(Q, R) = 1 - \frac{1}{N}\sum_{n=1}^{N} |dec(A_n) - conf_M(A_n) |,$$9

and intrinsic confidence is

$dec(A_n)$0

RLMF uses the same black-box style of consistency estimation for sentence-level intrinsic confidence, again with $dec(A_n)$1 sampled responses and NLI-style judgments (Liu et al., 30 May 2025, Liu et al., 30 Jun 2026).

Expressed confidence is likewise treated as a measured quantity rather than a given variable. MetaFaith extracts assertions and scores each assertion’s decisiveness with an LLM judge on a $dec(A_n)$2 scale, where 0 corresponds to hesitant, uncertain, non-committal language and 1 to confident, decisive language. The same general approach appears in the LRM paper, but there the unit is an entire reasoning step rather than an extracted factual assertion, because the target is the reasoning process itself rather than only the factual content embedded in it (Liu et al., 30 May 2025, Gani et al., 2 Jun 2026).

The LRM framework widens the notion of intrinsic confidence by introducing three estimator families. RCC (Recurrent Confidence Chain) combines final-layer hidden states and generated token probabilities through inter-step attention-like links and a recurrence

$dec(A_n)$3

with $dec(A_n)$4. DeepConf instead uses top-$dec(A_n)$5 next-token probabilities,

$dec(A_n)$6

followed by normalization to $dec(A_n)$7. The third estimator is sampling consistency, where for each step the model conditions on the original prompt and preceding steps, samples $dec(A_n)$8 continuations, and defines

$dec(A_n)$9

A central methodological result of that paper is that these estimators yield materially different FC judgments on the same traces (Gani et al., 2 Jun 2026).

RLMF adds an intermediate case in which the expressed signal is numerical rather than verbal. In Stage 1, the model outputs sentence-confidence pairs $conf_M(A_n)$0, and faithfulness is optimized directly between self-reported confidence $conf_M(A_n)$1 and intrinsic confidence estimate $conf_M(A_n)$2. Only after numerical FC is improved does Stage 2 map those scores into context-adaptable linguistic hedging via targeted rewriting (Liu et al., 30 Jun 2026).

4. Methods for improving faithful calibration

The first systematic study in this line is MetaFaith, a prompt-based, black-box, task-agnostic intervention designed to improve FC without fine-tuning or access to weights. Its best-performing strategy, MetSens+Hedge, combines metacognitive framing with a graded hedge lexicon such as “almost certain,” “likely,” “about even,” and “little chance.” The paper reports up to 61% improvement in faithfulness and an 83% human win rate over original generations, while also showing that standard prompt approaches provide only marginal gains and that factuality-based calibration techniques such as TS, FaR, and SAR can harm FC (Liu et al., 30 May 2025).

RLMF turns FC into a reinforcement-learning objective. For each completion $conf_M(A_n)$3, the overall reward combines a faithfulness term, a factual-calibration term, a correctness term, and two format terms. The primary FC reward is

$conf_M(A_n)$4

and the paper sets

$conf_M(A_n)$5

Its distinctive element is metacognitive advantage scaling: the model is asked to predict how faithful its own reported confidences were, and that self-judgment is used to refine completion rankings only when faithfulness is already above the group average. The reported outcome is state-of-the-art FC with up to 63% improvement over standard RL while preserving accuracy (Liu et al., 30 Jun 2026).

The same paper pairs RLMF with metacognitive data selection. For each training example, the model scores from 0 to 100 how well the linguistic decisiveness of its answer matches its internal confidence, and the final training set takes half the examples with the highest self-ratings and half with the lowest. The intended effect is to retain both metacognitively coherent examples and examples the model itself recognizes as mismatched or difficult. On the reported setup, this selection strategy outperforms both random selection and an active-learning-style baseline (Liu et al., 30 Jun 2026).

A complementary, directly relevant line is Retrieval-Augmented Linguistic Calibration (RALC). That work models linguistic confidence not as a scalar but as a Beta distribution

$conf_M(A_n)$6

introduces Faithfulness Divergence

$conf_M(A_n)$7

and applies Platt scaling to the mean

$conf_M(A_n)$8

while preserving concentration. A retrieval-augmented rewriting stage then propagates the calibrated distributional confidence back into language. The paper reports in-domain faithfulness and calibration improvements up to 66% and 58%, respectively (Yeh et al., 19 May 2026).

5. Faithful calibration in large reasoning models

FC becomes substantially harder in LRMs because long chain-of-thought traces do not have stable atomic units, show inconsistent structure across samples, and exhibit strong conditional dependence between later uncertainty and earlier reasoning. The LRM measurement paper identifies four specific complications: no stable step boundaries, inconsistent step structure across traces, unequal semantic importance of steps, and conditional dependence throughout the trace. Its methodological response is to define FC over reasoning-step spans and to estimate intrinsic confidence relative to a fixed prefix rather than by unconditional resampling (Gani et al., 2 Jun 2026).

The core innovation is prefix-conditioned sampling. For a step $conf_M(A_n)$9, the model samples continuations from

$R=\{s_1,\ldots,s_L\}$0

and compares each sampled continuation with the original step $R=\{s_1,\ldots,s_L\}$1 for consistency. This is meant to control structural and conditional variation: the question becomes whether, given exactly the same preceding reasoning context, the next local reasoning move is stable. To limit cost, the framework evaluates at most $R=\{s_1,\ldots,s_L\}$2 steps per trace, always retaining the first and last step and uniformly subsampling the rest (Gani et al., 2 Jun 2026).

Empirically, the paper reports that cMFG$R=\{s_1,\ldots,s_L\}$3 for LRMs is typically in the 0.64–0.78 range, so the models are not catastrophically unfaithful but remain far from ideal. It also shows that reasoning does not automatically improve FC. In same-backbone comparisons, reasoning-tuned models often sound more cautious—more hesitation, self-questioning, and correction language—without a corresponding drop in internal confidence, which lowers faithfulness. The paper further finds that prompt interventions that help non-reasoning LLMs do not improve FC in the reasoning setting, and that different intrinsic-confidence estimators produce divergent assessments of the same traces (Gani et al., 2 Jun 2026).

A direct implication, stated in the paper, is that apparent deliberation should not be treated as evidence of faithful introspective confidence. More reasoning, more hedging, or higher answer accuracy do not by themselves guarantee better FC. In that sense, the LRM work reframes FC as a separate alignment and reliability target for chain-of-thought systems (Gani et al., 2 Jun 2026).

6. Adjacent literatures and related notions

Several adjacent literatures address closely related ideas without using FC in the exact LLM sense. FairCal studies subgroup-conditional calibration in face verification. It defines a model as fairly-calibrated when calibration holds conditionally on subgroup membership and uses this to reduce false positive rate disparities without requiring sensitive attributes at inference. Its relevance to FC is indirect but substantial: it shows that global calibration can be insufficient when the real target is conditional or subgroup-aware reliability (Salvador et al., 2021).

In robot perception, $R=\{s_1,\ldots,s_L\}$4-Cal addresses aleatoric uncertainty calibration rather than linguistic uncertainty. Its central claim is that predictive uncertainties often do not faithfully represent the true underlying uncertainties or process noise, and that calibration is a cross-sample distributional property rather than a per-sample one. The method enforces that canonicalized residuals match a target distribution such as $R=\{s_1,\ldots,s_L\}$5 or $R=\{s_1,\ldots,s_L\}$6, making it a strong analogue of faithfulness for regression-style uncertainty estimation (Bhatt et al., 2021).

In federated learning, FedCal separates local calibration from global calibration and proves that under heterogeneous client label marginals, naive FedAvg aggregation can have a nonzero asymptotic lower bound on global calibration error. Its target is top-label confidence calibration rather than FC in the LLM sense, but it broadens the broader “faithful reliability” agenda by showing that a model can appear acceptable locally and still be globally miscalibrated under client heterogeneity (Peng et al., 2024).

A different adjacency appears in Faithful Knowledge Distillation, where the target is teacher-relative local confidence alignment rather than label-relative calibration. There, a student is a faithful imitator of a teacher around $R=\{s_1,\ldots,s_L\}$7 if

$R=\{s_1,\ldots,s_L\}$8

and the paper introduces empirical and certified procedures for evaluating this faithfulness bound. This is a relational notion of calibration, but it is again closer to “confidence should faithfully reflect a reference internal state” than to ordinary empirical calibration (Lamb et al., 2023).

The acronym itself is also ambiguous. In “Feature Clipping for Uncertainty Calibration,” FC means Feature Clipping, a post-hoc method that clips penultimate-layer feature values to reduce overconfidence. That work is about conventional calibration, not Faithful Calibration, and should not be conflated with the LLM FC literature despite the shared acronym (Tao et al., 2024).

7. Limitations, controversies, and open questions

A central limitation across direct FC papers is that intrinsic confidence is proxied rather than observed. MetaFaith and RLMF use response consistency under resampling; the LRM paper adds hidden-state and token-probability estimators; RALC models audience-perceived confidence distributions. None of these is taken as ground truth for an internal belief state, and the LRM study explicitly shows that different estimators can disagree strongly on the same reasoning trace (Liu et al., 30 May 2025, Liu et al., 30 Jun 2026, Gani et al., 2 Jun 2026, Yeh et al., 19 May 2026).

A second limitation is measurement dependence on judges and audiences. Direct FC work typically scores linguistic decisiveness with LLM-as-a-judge prompts, while RALC uses an ensemble of evaluator models as reader surrogates and validates them against human annotations. This suggests practical usefulness, but it also implies that FC is partly audience-relative: the same hedge may induce different beliefs for different readers, domains, or cultures. RALC makes this explicit by modeling linguistic confidence as a distribution over plausible perceived probabilities rather than as a single number (Yeh et al., 19 May 2026).

A third limitation is that FC does not subsume factual verification. RLMF states explicitly that improved faithful uncertainty expression is not a substitute for factual verification. A model can faithfully communicate low confidence and still be wrong, or faithfully communicate high confidence in a stable but incorrect belief state. MetaFaith similarly shows that factual calibration techniques can worsen FC, which underscores that the two targets are distinct and sometimes in tension (Liu et al., 30 Jun 2026, Liu et al., 30 May 2025).

There are also task- and system-specific open problems. The LRM paper shows that FC over long reasoning traces is unusually fragile because of step segmentation, path dependence, and estimator disagreement. RLMF raises self-referential failure modes and reward-hacking concerns when self-judgments are used as RL signal, although its advantage-scaling design is intended to reduce this risk. MetaFaith and RLMF are English-centric, and RALC notes that linguistic uncertainty expression varies with context and audience. Taken together, these results suggest that FC is best understood not as a solved calibration subroutine, but as an active research area at the intersection of uncertainty estimation, metacognition, language generation, and human-facing reliability (Gani et al., 2 Jun 2026, Liu et al., 30 Jun 2026, Liu et al., 30 May 2025, Yeh et al., 19 May 2026).