Pre-Calibration Smoothing

• Pre-calibration smoothing is a design principle that pre-processes data—such as hard labels, raw scores, or measurements—with smoothing techniques to mitigate over-confidence and noise sensitivity.
• It is employed across diverse domains, including probabilistic classification (soft labeling), signal processing (spectral filtering), and financial modeling (local volatility denoising).
• Empirical studies report significant calibration error reductions, such as lowering ECE values by up to 68.5%, while also sometimes boosting task performance.

Pre-Calibration Smoothing denotes a family of procedures that deliberately smooth an object before a downstream calibration, estimation, or decision step. Across the literature, the smoothed object varies: hard training targets, raw classifier scores, OFF-source spectra, raw visibilities, implied-volatility quotes, emulator and discrepancy terms, state trajectories, or quantized weight distributions. The common motif is that smoothing is introduced upstream so that subsequent calibration is better behaved, less over-confident, less noise-sensitive, or less ill-posed. In some settings it acts directly inside the training loop; in others it is a preprocessing stage applied to measurements; and in still others it is built into the definition of the calibration functional itself (Wei et al., 2022, Yamaki et al., 2012, Lucena, 2018, Yang et al., 19 Sep 2025).

1. Conceptual scope and formal role

In probabilistic classification, pre-calibration smoothing is often implemented by replacing hard one-hot targets with soft targets whose entropy depends on example difficulty or model uncertainty. In the histopathology setting, this is explicitly described as a “pre-calibration” step that occurs inside ordinary training, “long before any post-hoc calibration like temperature scaling is applied,” with the smoothed targets used directly in cross-entropy (Wei et al., 2022).

In signal-processing and observational science, the same term refers to smoothing raw calibration references or raw measurements before fitting calibration parameters. Smoothed Bandpass Calibration in radio spectroscopy applies spectral smoothing only to the OFF-source spectrum before bandpass correction, while preserving the ON-source spectral resolution (Yamaki et al., 2012). In 21 cm interferometry, custom time-domain filters are applied to visibilities prior to calibration so that mutual-coupling artifacts do not force gains to absorb unsmooth chromatic structure (Charles et al., 2024). In local-volatility calibration, a fully automatic local-regression stage denoises implied volatilities or prices before any off-the-shelf Dupire inversion is run (Yang et al., 19 Sep 2025).

A distinct but related line of work smooths the calibration functional itself. SmoothECE replaces binned reliability-diagram constructions with kernel smoothing of prediction–outcome pairs before computing a calibration error, and smooth calibration replaces discontinuous exact-bin tests by Lipschitz test functions or kernels (Błasiok et al., 2023, Foster et al., 2022, Gopalan et al., 16 Mar 2026). This suggests that “pre-calibration smoothing” is not tied to a single mathematical object; rather, it names a design principle in which high-frequency or discontinuous structure is softened before calibration is assessed or enforced.

2. Target smoothing in predictive models

A prominent form of pre-calibration smoothing replaces hard labels with sample-dependent soft targets. In histopathology image classification, one starts from one-hot labels $y_i\in\mathbb{R}^K$ and defines agreement-aware smoothing using annotator agreement

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$

with

$$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$

If annotator agreement is unavailable, confidence-aware smoothing substitutes a baseline model confidence $c_i$: $$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$ Training then replaces each hard target in the cross-entropy loss by $y_i^{\text{smooth}}$. On the MHIST colorectal-polyp task, the baseline model’s ECE was $8.9\%$; with piecewise or nonlinear agreement-aware smoothing this fell to about $2.8\%$, a $68.5\%$ relative reduction, and AUC rose from $84.7\%$ to $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$0. Confidence-aware smoothing achieved an ECE drop from $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$1 to $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$2 and up to $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$3 AUC over baseline (Wei et al., 2022).

In span-based neural NER, boundary smoothing reassigns probability mass from annotated spans to neighboring spans. For a sentence of length $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$4, candidate spans are

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$5

and the neighboring set within Manhattan radius $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$6 is

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$7

With smoothing weight $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$8,

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$9

This is applied directly to training targets, before any post-hoc calibration. Reported ECE values fell from $$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$0 to $$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$1 on CoNLL-2003 and from $$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$2 to $$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$3 on OntoNotes-5. The same work reports flatter loss landscapes, flatter neural minima, and a shift from aggressive over-confidence toward slight under-confidence when $$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$4 is too large (Zhu et al., 2022).

Adaptive label smoothing with self-knowledge makes the smoothing parameter itself instance-specific. For each example $$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$5, predictive entropy is

$$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$6

and the smoothing weight is

$$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$7

The prior distribution is taken from a past checkpoint $$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$8 selected on a held-out validation set by BLEU, with

$$y_i^{\text{smooth}} = (1 - \alpha a_i)\,y_i \;+\;\frac{\alpha a_i}{K}\,\mathbf{1}, \qquad \alpha\in(0,1].$$9

The final loss is

$c_i$0

On Multi30K and IWSLT14, ECE reportedly dropped from $c_i$1 and $c_i$2, with similar decreases in MCE (Lee et al., 2022).

Fixed label smoothing in pre-trained Transformers uses

$c_i$3

and optimizes

$c_i$4

The reported pattern is asymmetric across regimes: label smoothing lowers out-of-domain ECE, but standard MLE plus post-hoc temperature scaling yields the lowest in-domain ECE. Representative values include BERT on SNLI $c_i$5 for MLE versus $c_i$6 for LS in-domain$c_i$7 and MNLI $c_i$8 for MLE versus $c_i$9 for LS out-of-domain$$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$0, and RoBERTa on SWAG $$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$1 for MLE versus $$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$2 for LS in-domain$$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$3 and HellaSWAG $$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$4 for MLE versus $$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$5 for LS out-of-domain$$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$6 (Desai et al., 2020).

3. Score-space smoothing and smooth calibration functionals

Another major branch of pre-calibration smoothing operates on uncalibrated scores rather than labels. SplineCalib fits a non-parametric calibration map $$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$7 using a natural cubic spline basis

$$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$8

and minimizes penalized log-loss

$$y_i^{\text{smooth}} = (1 - \alpha c_i)\,y_i \;+\;\frac{\alpha c_i}{K}\,\mathbf{1}.$$9

In basis form this becomes logistic regression with quadratic roughness penalty $y_i^{\text{smooth}}$0. For over-confident models, a bounded “compact logit” transform $y_i^{\text{smooth}}$1 is applied before fitting. Multi-class calibration is handled one-vs-all followed by renormalization, and a cross-validated variant reuses all training data. Reported results include MIMIC-RF log-loss $y_i^{\text{smooth}}$2, Adult $y_i^{\text{smooth}}$3 with compact logit, and CIFAR-10 CV-SplineCalib log-loss $y_i^{\text{smooth}}$4 with accuracy $y_i^{\text{smooth}}$5 (Lucena, 2018).

SmoothECE smooths the prediction–outcome relation itself with a reflected Gaussian kernel

$y_i^{\text{smooth}}$6

and estimates the conditional mean by

$y_i^{\text{smooth}}$7

The smoothed residual and smoothed prediction density are

$y_i^{\text{smooth}}$8

and the scale-$y_i^{\text{smooth}}$9 calibration error is

$8.9\%$0

The bandwidth is chosen by the fixed-point rule

$8.9\%$1

which the authors state is unique because $8.9\%$2 is continuous and monotonically decreasing. The resulting measure is presented as hyperparameter-free, consistent, and sample-efficient, and is implemented in the relplot package (Błasiok et al., 2023).

Smooth calibration in the forecasting literature replaces discontinuous calibration tests by smooth ones. One formulation defines

$8.9\%$3

where $8.9\%$4 is $8.9\%$5-Lipschitz. The same work states

$8.9\%$6

and studies a smoothing mechanism that adds uniform noise $8.9\%$7 to predictions and clips back to $8.9\%$8,

$8.9\%$9

It further characterizes smooth calibration in terms of earth-mover distance to the closest perfectly calibrated distribution, up to constant factors (Gopalan et al., 16 Mar 2026). A related framework replaces the exact-bin indicator $2.8\%$0 by a Lipschitz kernel $2.8\%$1, yielding deterministic, finite-recall, stationary, grid-valued forecasting procedures that guarantee smooth calibration even when forecasts are leaked (Foster et al., 2022).

4. Signal and measurement preprocessing before calibration

In radio spectroscopy, Smoothed Bandpass Calibration was introduced to reduce the thermal-noise contribution of the OFF-source blank-sky measurement used for bandpass calibration. The core step is to smooth only the OFF-source spectrum: $2.8\%$2 with $2.8\%$3. The calibrated spectrum is then

$2.8\%$4

The smoothing window is selected by the bottom of the Spectral Allan Variance curve, while the usable ON–OFF duty cycle is selected by Time-based Allan Variance. For the targeted noise level of $2.8\%$5 as a ratio to the system noise, the optimal smoothing window was $2.8\%$6 channels in a $2.8\%$7-channel bandwidth, and the optimal scan pattern was $2.8\%$8 s ON $2.8\%$9 $68.5\%$0 s OFF. The noise level was reduced by a factor of $68.5\%$1, the required telescope time fell from $68.5\%$2 s to $68.5\%$3 s, and the calculated efficiency improvement was $68.5\%$4, $68.5\%$5, and $68.5\%$6 for single-beam, dual-beam, and OTF observations, respectively (Yamaki et al., 2012).

In low-frequency 21 cm cosmology, pre-calibration smoothing is implemented as time-axis filtering of raw visibilities before redundant and absolute calibration. For a visibility time series $68.5\%$7, one defines the fringe-rate transform

$68.5\%$8

and applies a filter transfer function $68.5\%$9. Two filters are described. The notch filter sets

$84.7\%$0

with $84.7\%$1 mHz. The main-lobe filter uses a Gaussian fit $84.7\%$2 and sets $84.7\%$3 for $84.7\%$4 and $84.7\%$5 outside. After inverse transformation, calibration is run on $84.7\%$6 rather than $84.7\%$7, but the recovered gains are then applied to the unfiltered data. Reported gains are much smoother, with reduced power at delays $84.7\%$8 ns. Foreground leakage at $84.7\%$9 is suppressed by $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$00, residuals approach the thermal noise at $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$01 in the low band and by $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$02 in the high band, and the dynamic range at $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$03 ns improves by a factor of $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$04 (Charles et al., 2024).

These examples illustrate a recurrent pattern: smoothing is applied only to the part of the measurement chain that carries nuisance structure. In SBC, only OFF-source data are smoothed; in HERA calibration, filtered data are used to estimate gains, but the final calibrated product is formed from unfiltered visibilities. This suggests that pre-calibration smoothing need not imply a blanket loss of resolution.

5. Smoothing as a precursor to ill-posed and high-dimensional calibration

In local-volatility calibration, pre-calibration smoothing appears as a fully automatic denoising stage for market observables. For a fixed maturity $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$05, noisy implied volatilities satisfy

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$06

and at target strike $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$07 one fits a local polynomial of order $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$08 by weighted least squares: $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$09 The smoothed value is $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$10, and $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$11 are selected by minimizing asymptotic conditional mean squared error. Stage 2 then feeds the smoothed implied-volatility grid into any standard LV calibration, such as finite-difference Dupire inversion. In an SVI-simulated market, absolute calibration error over three moneyness buckets fell from $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$12 to $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$13; for a real-world AAPL pre-earnings “W-shaped” IV curve, the proposed smoothing plus standard calibration yielded a smooth LV surface, kept model-IV fit within spreads, achieved fail ratio $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$14, and kept absolute IV-error by moneyness/maturity bucket below $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$15 (Yang et al., 19 Sep 2025).

Bayesian Smoothing Spline ANOVA uses smoothing splines for both computer-model emulation and model–reality discrepancy in calibration problems with categorical parameters and correlated outputs. The emulator and discrepancy admit basis expansions such as

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$16

with Gaussian priors on coefficients and component-specific smoothing parameters $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$17 or covariance matrices $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$18. The resulting framework supports Gibbs updates for smoothing parameters, handles categorical inputs via a discrete-ANOVA covariance, and scales linearly in $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$19 once design matrices are built. The same summary explicitly notes that pre-calibration smoothing can be implemented by projecting raw data onto the same basis and minimizing a penalized least-squares objective before calibration (Storlie et al., 2014).

A different use of pre-calibration smoothing appears in dynamical systems, where noisy observations are smoothed by soft adherence to governing equations. With measurements $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$20 and dynamics $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$21, one solves

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$22

where $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$23 penalizes deviations from a Runge–Kutta update and $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$24 penalizes data mismatch. The hyperparameter $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$25 balances adherence to dynamics and fidelity to observations, with reported robust values $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$26 to $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$27. In Lorenz-96 with true $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$28, the recovered parameter was $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$29, and across Lorenz-63, Lorenz-96, Kuramoto–Sivashinsky, and nonlinear Schrödinger examples the method outperformed Ensemble RTS smoothing when dynamics were known (Rudy et al., 2018).

In post-training quantization, statistical pre-calibration is formulated as matching the quantized weight distribution $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$30 to the original distribution $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$31 by minimizing

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$32

with

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$33

A local Taylor expansion under pseudo-activations $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$34 yields the soft-thresholding rule

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$35

used to classify weights as salient or common before separate quantization. The method is described as a precursor to calibration-based PTQ methods and is reported to match or exceed leading calibration baselines on perplexity, zero-shot reasoning, and coding tasks, while quantization on LLaMA-7B took approximately $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$36 s versus AWQ’s $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$37 s and SpQR’s $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$38 s (Ghaffari et al., 15 Jan 2025).

6. Calibration metrics, interpretation, and trade-offs

The most common metric in the classification papers is Expected Calibration Error. With confidence bins $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$39,

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$40

and

$$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$41

This binning-based definition underlies the histopathology and Transformer studies, while SmoothECE was introduced precisely because binning and ordinary ECE suffer from discontinuity (Wei et al., 2022, Błasiok et al., 2023).

A recurrent empirical finding is that smoothing can improve calibration without sacrificing, and sometimes while improving, task performance. Histopathology label smoothing improved both ECE and AUC; SplineCalib improved log-loss and occasionally accuracy; adaptive label smoothing in NLG improved BLEU alongside ECE and MCE; and local-regression smoothing in LV calibration preserved fit to market observables while stabilizing Greeks (Lucena, 2018, Lee et al., 2022, Yang et al., 19 Sep 2025). A plausible implication is that smoothing often acts simultaneously as calibration control and regularization.

The reported trade-offs are equally important. Label smoothing is not uniformly beneficial in every regime: in pre-trained Transformers it improves out-of-domain calibration but can worsen in-domain ECE relative to MLE, especially when compared with post-hoc temperature scaling (Desai et al., 2020). In NER, excessively large $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$42 shifts the model toward slight under-confidence (Zhu et al., 2022). In HERA calibration, notch filtering suppresses coupling more effectively but reduces sky signal and raises the noise floor in recovered gains, whereas main-lobe filtering preserves more sky power but permits residual coupling (Charles et al., 2024). In local-volatility calibration, regularization that directly penalizes $$a_i := \text{fraction of annotators who chose the majority class for } x_i,$$43 can smooth the LV surface but may drive model IV outside market quotes unless the penalty is tiny, whereas preprocessing by automatic local regression isolates the statistical bias–variance trade-off from the subsequent arbitrage-constrained calibration (Yang et al., 19 Sep 2025).

The literature therefore distinguishes several non-equivalent meanings of smoothing. Some methods smooth targets before fitting a predictor; some smooth scores before mapping them to calibrated probabilities; some smooth measurements before estimating calibration parameters; some smooth the test of calibration itself. What unifies them is the replacement of a high-variance, discontinuous, or noise-amplifying object by a controlled, data-dependent surrogate before the calibration stage is executed or evaluated.