Prediction-Powered Inference Overview

• PPI is a semi-supervised framework that uses machine predictions as low-variance surrogates to augment limited high-quality labels for unbiased estimation of population parameters.
• It employs a rectification mechanism to estimate and remove prediction bias, ensuring valid confidence intervals even when using imperfect prediction models.
• Extended methods like PPI++ optimize efficiency via adaptive power tuning, balancing bias and variance to improve finite-sample performance.

Prediction-powered inference (PPI) is a semi-supervised inferential framework for combining a small set of expensive, high-quality gold-standard labels with a large set of cheap machine predictions or pseudo-labels in order to estimate population quantities and construct valid uncertainty statements. In its canonical form, PPI uses predictions on the unlabeled sample as a low-variance surrogate and then estimates and removes prediction bias using the labeled sample, so that validity does not depend on the correctness of the prediction model. The framework was introduced for means, quantiles, and regression coefficients, and was later extended by PPI++ through adaptive power tuning, which reweights the prediction-based correction to improve efficiency and computability (Angelopoulos et al., 2023, Angelopoulos et al., 2023).

1. Core formulation

The basic PPI setting consists of a labeled sample of size $n$, $\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$, an unlabeled sample of size $N$, $\{X_j\}_{j=1}^N$, and a fixed prediction rule $f$ that produces pseudo-labels $F=f(X)$. The archetypal target is the population mean

$\theta \coloneqq \mathbb E[Y].$

With $F_i=f(X_i)$ on labeled points and $F_j=f(X_j)$ on unlabeled points, the PPI estimator for the mean is

$$\hat{\theta}_{\text{PPI}} = \frac{1}{n}\sum_{i \in \mathcal{D}_n} Y_i + \frac{1}{N}\sum_{j \in \mathcal{D}'_N} F_j - \frac{1}{n}\sum_{i \in \mathcal{D}_n} F_i,$$

which is unbiased because the prediction mean enters once on the unlabeled sample and once, with opposite sign, on the labeled sample (Mani et al., 26 May 2025).

PPI++ introduces a scalar interpolation parameter $\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$0 and replaces full correction by

$\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$1

Here $\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$2 recovers the labeled-only estimator, $\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$3 recovers base PPI, and intermediate values trade off classical and prediction-assisted estimation. When population moments are known, the variance-optimal weight is

$\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$4

with $\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$5 and $\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$6; in the effectively infinite-$\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$7 regime this simplifies to $\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$8 (Mani et al., 26 May 2025).

The same logic extends beyond mean estimation. In the original loss-based formulation, the parameter of interest is defined by

$\mathcal D_n=\{(X_i,Y_i)\}_{i=1}^n$9

and PPI constructs a rectified empirical loss

$N$0

where $N$1 is the supervised empirical loss on labeled data, $N$2 is the pseudo-label loss on unlabeled data, and $N$3 is the pseudo-label loss on labeled data. Because the two pseudo-label terms have the same expectation under the shared covariate distribution, the rectified loss is unbiased for the population risk. This supports means, quantiles, linear regression coefficients, logistic regression coefficients, and more general M-estimators (Angelopoulos et al., 2023, Shoham et al., 26 Oct 2025).

2. Asymptotic theory and efficiency

Original PPI established valid confidence intervals for quantities such as means, quantiles, and linear and logistic regression coefficients, without making assumptions on the machine-learning algorithm that supplies the predictions, while showing that more accurate predictions translate to smaller confidence intervals (Angelopoulos et al., 2023). PPI++ recast the method as a computationally lightweight M-estimation procedure: rather than testing a rectified score over a grid of candidate parameters, it minimizes a rectified loss and then applies standard asymptotic normality and sandwich covariance estimation. In this formulation, the asymptotic covariance depends on the Hessian of the target loss and on two variance components, one from the prediction-based term and one from the rectifier, and power tuning chooses $N$4 to minimize a scalarization of that covariance (Angelopoulos et al., 2023).

A later semiparametric analysis formalized PPI as an M-estimation problem under simple random sampling without replacement and derived its efficient influence function. In that framework, PPI is consistent and asymptotically normal, and it attains the semiparametric efficiency lower bound when the predictor is score-calibrated, meaning that $N$5 equals $N$6 almost surely. For mean estimation, score calibration reduces to

$N$7

The same work also developed asymptotic theory for learned predictors: cross-fitted PPI for semiparametric mean estimation requires only $N$8-consistency of the learned predictor, and a single-fit variant with variance correction is available for linear smoothers, notably kernel ridge regression with an unpenalized intercept (Lee et al., 7 Jun 2026).

This asymptotic picture clarifies a central feature of the framework. PPI and PPI++ do not treat predictions as substitutes for gold-standard outcomes; rather, they use predictions as auxiliary variables inside an unbiased or asymptotically unbiased estimator. Efficiency depends on how well the prediction rule aligns with the target score or residual structure, not on predictive accuracy in isolation. This suggests why later work could simultaneously prove semiparametric optimality under score calibration and still identify finite-sample regimes in which prediction assistance harms precision.

3. Finite-sample behavior and the “no free lunch” phenomenon

The best-known asymptotic result for PPI++ is a “free lunch” statement: the asymptotic variance of PPI++ is never larger than the variance obtained from using gold-standard labels alone, regardless of the quality of the pseudo-labels (Angelopoulos et al., 2023). Exact finite-sample analysis shows that this statement does not extend verbatim to the small-$N$9 regime in which PPI is often most attractive. For the mean estimation problem, split-sample PPI++ satisfies

$\{X_j\}_{j=1}^N$0

The three terms are, respectively, the classical variance, the efficiency gain from using the true covariance, and the efficiency loss from estimating that covariance. Hence split-sample PPI++ has worse MSE than the classical estimator iff

$\{X_j\}_{j=1}^N$1

In words, PPI++ helps only if the squared signal exceeds the MSE of the covariance estimator (Mani et al., 26 May 2025).

This finite-sample theory yields explicit thresholds. For Gaussian data, cross-fit PPI++ improves over the classical estimator iff the correlation between pseudo- and gold-standard labels satisfies

$\{X_j\}_{j=1}^N$2

and a similar $\{X_j\}_{j=1}^N$3 threshold holds for binary labels. In the pure no-signal case of independent Gaussian $\{X_j\}_{j=1}^N$4 and $\{X_j\}_{j=1}^N$5, split-sample PPI++ has

$\{X_j\}_{j=1}^N$6

so it is strictly worse than using gold-standard labels alone (Mani et al., 26 May 2025).

The same analysis also distinguishes single-sample and split-sample implementations. Single-sample PPI++ reuses the same labeled data to estimate $\{X_j\}_{j=1}^N$7 and $\{X_j\}_{j=1}^N$8, and is biased in finite samples: $\{X_j\}_{j=1}^N$9 Its naive plug-in variance estimator is always smaller than the classical plug-in variance, even in regimes where the true MSE is larger, so intervals can be too narrow and may severely undercover. Split-sample PPI++ is unbiased in finite samples, and cross-fitting averages two split estimators at the cost of an effective sample-size penalty (Mani et al., 26 May 2025).

A related critique in linear regression reaches a similar conclusion from a different angle. In that setting, the PPI estimator is one member of a broader augmented class, and a weighted augmentation due to Chen and Chen can be asymptotically at least as efficient as using only the labeled data, whereas plain PPI can be less efficient than ignoring predictions. The proposed Chen–Chen estimator retains the robustness property that validity does not depend on the accuracy of the ML predictor (Gronsbell et al., 2024). Taken together, these results imply that “valid regardless of ML accuracy” should not be conflated with “uniform finite-sample efficiency.”

4. Survey-sampling roots, inferential modes, and informative labeling

For mean estimation, the PPI estimator is algebraically equivalent to the difference estimator of Cassel et al. (1976), and PPI++ corresponds to the generalized regression (GREG) estimator of Särndal et al. (2003). In this interpretation, the prediction $f$0 is simply an auxiliary variable, the labeled subset is a probability sample, and the rectifier is the sample mean of residuals $f$1. This equivalence does not require correctness of the prediction model; it requires only that labels and predictions arise from the same underlying finite population or superpopulation (Mozer, 19 Mar 2026).

The equivalence also clarifies an important distinction in mode of inference. In the design-based survey view, the finite population is fixed, the estimand is the finite-population mean or SATE, and randomness comes from the sampling design. In the superpopulation view common in the PPI literature, $f$2 are i.i.d. draws from a distribution, the estimand is $f$3 or PATE, and randomness comes from sampling from that distribution. The point estimator is the same, but the interpretation of coverage and the variance decomposition differ (Mozer, 19 Mar 2026).

This perspective is especially consequential for subgroup and causal estimands. If treatment effects are estimated using a pooled rectifier across treatment arms, and arm-specific mean prediction errors are $f$4 and $f$5, then the resulting treatment-effect estimator has bias

$f$6

That bias does not shrink with more labels. The remedy is arm-specific or subgroup-specific rectifiers, which separately estimate and remove the prediction bias within each subgroup (Mozer, 19 Mar 2026).

Informative labeling introduces a further departure from the default PPI assumptions. Standard PPI implicitly assumes simple random sampling or labels missing completely at random. When labeling probabilities vary by covariates, $f$7, the unweighted rectifier is biased for the population mean residual. An inverse-probability-weighted extension replaces the rectifier by Horvitz–Thompson or Hájek estimators of the residual mean: $f$8 leading to

$f$9

Under MAR-type informative labeling and correct inclusion probabilities, the HT version is design-unbiased and the Hájek version is approximately unbiased with typically lower variance. Standard PPI under MCAR is already a Hájek ratio estimator applied to residuals (Datta et al., 13 Aug 2025).

5. Methodological variants and generalizations

The PPI literature has expanded rapidly beyond the original single-predictor, single-estimand formulation. Several variants retain the same core motif—prediction-assisted point estimation plus a labeled-data rectifier—while changing the inferential setting, the estimator family, or the optimization target.

Variant	Core mechanism	Notable property
Bayesian PPI (Hofer et al., 2024)	Monte Carlo integration over posterior means and proportions	Enables task-appropriate PPI methods for discrete autoraters and nonlinear predictor–human relationships
Federated PPI (Luo et al., 2024)	Train local models, aggregate through federated learning, and compute PPI from summaries	Produces statistically valid conclusions without sharing private information
Prediction-Powered Adaptive Shrinkage (Li et al., 20 Feb 2025)	Debias predictions within each task, then shrink multiple means toward prediction-based targets	CURE-tuned estimator is asymptotically optimal in Bayes MSE over its shrinkage family
PPI with IPW (Datta et al., 13 Aug 2025)	Replace the unweighted rectifier by HT or Hájek residual weighting	Remains valid under informative labeling
Prediction-Powered SSL with online power tuning (Shoham et al., 26 Oct 2025)	Use prediction-powered gradients and update $F=f(X)$0 by one-dimensional online learning	Gives an unbiased stochastic gradient and improved performance over classic SSL baselines and offline-tuned PPI methods
MOE-powered inference (Gu et al., 30 Apr 2026)	Choose a mixture of experts that minimizes the variance of the PPI estimator	Adapts to unknown expert quality and enjoys a best-expert guarantee

Bayesian PPI retains the difference-estimator intuition but replaces closed-form variance derivations by posterior sampling for intermediate quantities such as means, proportions, and multinomial probabilities. This makes it straightforward to construct PPI procedures for abstaining LLM judges, side-by-side evaluation, and stratified nonlinear corrections (Hofer et al., 2024). Federated PPI moves the same logic into decentralized settings: local prediction-based and rectifier statistics are computed on each client and aggregated with weights that reproduce the centralized empirical average, enabling inference without sharing raw data (Luo et al., 2024).

Other extensions broaden the inferential target. Prediction-Powered Adaptive Shrinkage combines per-task power-tuned PPI with empirical Bayes shrinkage across many related means, using the predictions both to debias each task and as task-specific shrinkage targets (Li et al., 20 Feb 2025). Prediction-powered semi-supervised learning applies the same rectification principle to gradients rather than parameters, defining

$F=f(X)$1

which is unbiased for the true population gradient for any $F=f(X)$2, and then tuning $F=f(X)$3 online by minimizing cumulative gradient second moments (Shoham et al., 26 Oct 2025). MOE-powered inference replaces a single predictor by a linear or more general mixture of experts and selects the mixture that minimizes the variance of the residual score, with non-asymptotic coverage-error bounds for the resulting confidence sets (Gu et al., 30 Apr 2026).

These variants suggest that PPI is less a single estimator than a design pattern: construct a prediction-based surrogate for the target estimating equation or loss, estimate its discrepancy from the gold-standard version on labeled data, and choose the prediction contribution to optimize variance, robustness, or both.

6. Applications, power analysis, and practical use

PPI was introduced with applications in proteomics, astronomy, genomics, remote sensing, census analysis, and ecology, reflecting its natural fit for scientific settings in which human or experimental labeling is expensive but prediction models can be applied at scale (Angelopoulos et al., 2023). Later studies expanded this range to biological sequences and galaxies (Mani et al., 26 May 2025), Amazon review ratings and Galaxy Zoo subgroup estimation (Li et al., 20 Feb 2025), NHANES BMI under informative labeling (Datta et al., 13 Aug 2025), UCI Energy Efficiency data (Lee et al., 7 Jun 2026), and evaluation tasks involving LLM judges and side-by-side model comparisons (Hofer et al., 2024).

A design-stage perspective is now available. Closed-form power and sample-size formulas have been derived for one-sample means, two-sample comparisons, paired means, $F=f(X)$4 relative risks and odds ratios, and linear and logistic regression contrasts. In the many-predictions regime, a useful rule of thumb is that the reduction in required labeled samples relative to classical designs scales roughly with the $F=f(X)$5 between the predictions and the ground truth. For the one-sample mean,

$F=f(X)$6

when $F=f(X)$7, so a predictor with $F=f(X)$8 cuts the labeled sample size roughly in half, whereas $F=f(X)$9 leaves only about $\theta \coloneqq \mathbb E[Y].$0 of the classical label requirement (Chen et al., 17 Mar 2026).

The same work also shows why discriminative performance metrics can mislead when translated into inferential gains. In dermoscopy melanoma detection, AUROCs of 0.72, 0.79, and 0.83 still led to only minimal label savings because the prevalence was about $\theta \coloneqq \mathbb E[Y].$1, so $\theta \coloneqq \mathbb E[Y].$2 remained small on the outcome scale (Chen et al., 17 Mar 2026). This suggests that PPI planning should be based on residual variance or score-level correlation, not on ranking metrics alone.

Several practical cautions recur across the literature. When the predictor is learned on the same data used for inference, cross-fitting or explicit variance correction is needed for reliable coverage (Lee et al., 7 Jun 2026). When subgroup prediction errors differ, pooled rectifiers can induce systematic bias, so subgroup-specific rectifiers should be standard for subgroup means and treatment effects (Mozer, 19 Mar 2026). When labeling is informative rather than MCAR, inverse-probability weighting should replace the unweighted rectifier (Datta et al., 13 Aug 2025). And when labeled sample sizes are small, finite-sample thresholds on pseudo-label correlation can dominate the asymptotic “free lunch,” so classical estimators may remain preferable until the signal-to-estimation-noise ratio crosses the relevant $\theta \coloneqq \mathbb E[Y].$3 boundary (Mani et al., 26 May 2025).

In this sense, prediction-powered inference occupies an intermediate position between classical design-based estimation and modern ML-assisted analysis. It preserves inferential validity under weak assumptions, can attain semiparametric efficiency under score calibration, and supports a broad family of extensions; yet its actual gains remain governed by prediction quality, labeling design, subgroup error structure, and finite-sample covariance estimation.