General Synthetic-Powered Inference (GESPI)

• GESPI is a statistical framework that integrates abundant synthetic signals with a small trusted dataset to enhance inferential efficiency while preserving validity.
• It employs techniques like synthetic-residual decomposition, power tuning, and guardrailed aggregation to correct bias and align synthetic with real data.
• Empirical studies show that GESPI notably reduces confidence interval widths and testing errors in fields such as genomics, astronomy, and ecology.

General Synthetic-Powered Inference (GESPI) denotes a family of statistical inference recipes in which abundant synthetic information—predictions, imputed labels, pseudo-labeled samples, proxy measurements, or generative-model outputs—is combined with a smaller trusted labeled sample to improve efficiency while preserving formal validity guarantees. Across the literature, GESPI is presented not as a single estimator but as a blueprint, recipe, or wrapper: synthetic information enlarges the effective information set, while a correction mechanism keeps inference anchored to the real-data target through rectification, debiasing, importance weighting, conformal calibration, empirical-likelihood constraints, or guardrail aggregation (Angelopoulos et al., 2023, Farzaneh et al., 4 Sep 2025, Bashari et al., 24 Sep 2025).

1. Definition and scope

GESPI is formulated for settings in which real labels are scarce, expensive, or imbalanced, but synthetic signals are plentiful. The synthetic component may be a predictor trained on independent data, an autorater score, a pseudo-label from a cross-fitted model, a counterfactual generator, a generative-model output, or a representation extracted from a pre-trained generative model. The inferential target may be a population mean, a quantile, regression coefficients, a prediction set, a counterfactual interval, a causal estimand, a confidence sequence, or a multiple-testing rejection set (Song et al., 28 Jan 2026, Imai et al., 5 Jul 2025).

This breadth is visible in the main lineages associated with the term. Prediction-Powered Inference (PPI) treats model predictions on a large unlabeled sample as a low-variance plug-in estimate and corrects its bias with labeled residuals (Angelopoulos et al., 2023). Synthetic data-powered conformal counterfactual inference augments scarce counterfactual calibration data with synthetic counterfactual labels and then debiases the induced risk estimate (Farzaneh et al., 4 Sep 2025). Synthetic-Powered Predictive Inference (SPI/SPPI) uses synthetic labeled calibration data to improve conformal prediction through score transport (Bashari et al., 19 May 2025). SynthBH uses pooled-data $p$-values but caps their influence so that false discovery rate control remains anchored to valid real-data $p$-values (Lee et al., 18 Feb 2026). A wrapper-style GESPI runs a base method on real data, relaxed real-data guardrails, and pooled real-plus-synthetic data, then combines the resulting actions by lattice operations (Bashari et al., 24 Sep 2025).

A recurring distinction in this literature is between synthetic information and trusted supervision. Synthetic information can be arbitrarily inaccurate, but validity is preserved because the final inferential object is corrected or guarded using real labels. This is the central conceptual departure from imputation-only or pseudo-label-only procedures, which treat synthetic labels as truth and therefore can be biased (Angelopoulos et al., 2023, Song et al., 28 Jan 2026).

2. Core statistical mechanics

The most common GESPI mechanism is a synthetic–residual decomposition. In the PPI formulation, if the target parameter $\theta_*$ satisfies an estimating equation

$\mathbb{E}[\psi(Z;\theta_*)] = 0,$

and $\phi(X;\theta)$ approximates $\mathbb{E}[\psi(Z;\theta)\mid X]$, then GESPI solves

$$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$

where $P_U$ and $P_L$ are sample averages over unlabeled and labeled data, respectively (Angelopoulos et al., 2023). The first term is a synthetic plug-in estimated on the large sample; the second is a labeled residual correction that removes bias. For mean estimation this becomes

$$\hat{\mu}_{\mathrm{PPI}} = \frac{1}{n_U}\sum_{j=1}^{n_U}\hat{f}(X_j) + \frac{1}{n_L}\sum_{i=1}^{n_L}\big(Y_i-\hat{f}(X_i)\big),$$

with asymptotic variance equal to the sum of a synthetic variance term and a residual variance term (Angelopoulos et al., 2023).

A second mechanism is power tuning. PPI++ introduces a weighting parameter $p$0 in the rectified objective,

$p$1

so that the method interpolates between classical inference ($p$2) and standard PPI ($p$3). The plug-in choice of $p$4 minimizes asymptotic variance and ensures that the resulting intervals always improve on classical labeled-only intervals asymptotically (Angelopoulos et al., 2023).

A third mechanism is guardrailed aggregation. The wrapper-style GESPI paper defines a two-sided construction

$p$5

which deterministically sandwiches the synthetic-powered action between the nominal real-data action and a relaxed guardrail action (Bashari et al., 24 Sep 2025). This construction does not modify the base algorithm internally; it wraps around it.

A fourth mechanism is risk control. In SP-CCI, the target is the miscoverage risk

$p$6

and calibration selects the smallest widening $p$7 such that an RCPS-style upper confidence bound $p$8 is at most $p$9. Synthetic labels enter the estimator of $\theta_*$0, but a PPI-style correction term debiases them (Farzaneh et al., 4 Sep 2025).

3. Principal methodological families

The literature now contains several distinct GESPI families, each centered on a different inferential object.

Family	Synthetic input	Validity mechanism
PPI / PPI++	predictions on unlabeled data	labeled residual correction; power tuning
Bayesian PPI	autorater scores or labels	proxy estimands with conjugate posteriors and Monte Carlo integration
EPI	predictions as auxiliary moments	empirical likelihood under stacked supervised and auxiliary constraints
MPPI	multiple pseudo-labeled sources	rectified source risks and weight selection by minimizing confidence-region volume
SP-CCI	synthetic counterfactual labels	RCPS miscoverage control with PPI-style debiasing
SPI / SPPI	synthetic labeled calibration data	score transporter and worst-case conformal coverage bounds
SynthBH	pooled-data $\theta_*$1-values	guarded synthetic-powered $\theta_*$2-values and PRDS-based FDR control
Wrapper GESPI	pooled outputs of any base method	min/max guardrails at $\theta_*$3 and $\theta_*$4

PPI is the foundational exemplar. It covers means, quantiles, linear regression, logistic regression, and general convex M-estimation, with asymptotically valid confidence intervals and no requirement that the predictor be unbiased or consistent (Angelopoulos et al., 2023). PPI++ strengthens this template by replacing grid-based testing with efficient convex optimization and by introducing power tuning, which yields a computationally lightweight procedure for parameters of arbitrary dimension (Angelopoulos et al., 2023).

Bayesian PPI recasts the same synthetic-powered logic in terms of task-appropriate proxy estimands and weak conjugate posteriors. This is particularly useful for discrete autoraters, abstaining judges, and paired evaluations, where chain-rule estimators or stratified estimators can be more natural than a direct residual-correction formula (Hofer et al., 2024). Empirical Likelihood Meets Prediction-Powered Inference replaces the rectified estimating equation with stacked empirical-likelihood constraints, yielding an EPI estimator that has asymptotic variance no larger than the fully supervised estimator and attains the semiparametric efficiency bound when the auxiliary moments span the predictable component of the supervised score (Wang et al., 18 Dec 2025).

Later extensions broaden the source structure. Multi-Source Prediction-Powered Inference aggregates several pseudo-labeled datasets and estimates the aggregation weights by minimizing the asymptotic volume of the resulting confidence region in homogeneous, covariate-shift, and domain-shift settings (Li et al., 19 Jun 2026). Prediction-Powered Inference Across Many Tasks uses cross-task recalibration and within-task rectification to improve task-specific inference when each task has only a handful of labels, while proving that gains beyond power-tuned PPI require nonlinear structure in the proxy–ground-truth relationship (Emmenegger et al., 28 May 2026). Anytime-valid, Bayes-assisted, Prediction-Powered Inference carries the framework to the sequential setting through confidence sequences built from Ville’s inequality and the method of mixtures (Kilian et al., 23 May 2025).

4. Prediction, counterfactual, causal, and testing instantiations

A major branch of GESPI concerns prediction sets and interval-valued inference. Synthetic-Powered Predictive Inference introduces a score transporter that aligns synthetic and real nonconformity scores by empirical quantile mapping, with a conservative window-based construction that yields finite-sample coverage guarantees without any assumptions about the real and synthetic data distributions. When score distributions are aligned, the method yields tighter prediction sets than standard split conformal prediction (Bashari et al., 19 May 2025).

SP-CCI provides a counterfactual version of the same principle. Under strong ignorability and overlap, it targets marginal coverage for individual counterfactual outcomes,

$\theta_*$5

even when treatment imbalance leaves very few real treated calibration points. Synthetic treated labels are generated for control-arm covariates, then incorporated into a debiased estimator of miscoverage risk. The resulting interval uses

$\theta_*$6

and the paper proves an RCPS-style high-probability marginal coverage guarantee under correct or boundedly misspecified importance weights (Farzaneh et al., 4 Sep 2025).

GESPI has also been applied to unstructured data. GenAI-Powered Inference uses open-source LLMs and diffusion models to extract low-dimensional representations $\theta_*$7 from text and images, then performs predictive or causal inference with regression adjustment, propensity modeling, and doubly robust estimation on $\theta_*$8. The defining causal assumption is conditional ignorability,

$\theta_*$9

with overlap in the $\mathbb{E}[\psi(Z;\theta_*)] = 0,$0-space (Imai et al., 5 Jul 2025). The synthetic component in this formulation is not only generated data but also internal representations of generative models.

Multiple testing provides a distinct GESPI regime. SynthBH defines guarded synthetic-powered $\mathbb{E}[\psi(Z;\theta_*)] = 0,$1-values

$\mathbb{E}[\psi(Z;\theta_*)] = 0,$2

and uses a BH-style step-up rule with a rank-adaptive cap. The pooled-data $\mathbb{E}[\psi(Z;\theta_*)] = 0,$3-values $\mathbb{E}[\psi(Z;\theta_*)] = 0,$4 need not be valid under the null; finite-sample FDR control is derived instead from the validity of the real-data $\mathbb{E}[\psi(Z;\theta_*)] = 0,$5 and an extended PRDS condition with respect to those real $\mathbb{E}[\psi(Z;\theta_*)] = 0,$6-values (Lee et al., 18 Feb 2026). The wrapper-style GESPI paper generalizes this guarded logic beyond testing by showing how conformal prediction, conformal risk control, one-sided hypothesis tests, and family-wise error procedures can all be wrapped by nominal, pooled, and guardrail runs of the same base method (Bashari et al., 24 Sep 2025).

5. Empirical record

The empirical literature consistently reports efficiency gains when synthetic information is informative and the correction mechanism is properly implemented. In the original PPI study, case studies in proteomics, astronomy, genomics, remote sensing, census analysis, and ecology showed that PPI significantly reduced confidence-interval widths relative to classical inference while retaining coverage, and the reported sample-size reductions ranged from about $\mathbb{E}[\psi(Z;\theta_*)] = 0,$7 to nearly $\mathbb{E}[\psi(Z;\theta_*)] = 0,$8 depending on task (Angelopoulos et al., 2023).

Bayesian PPI reports interval-width reductions of $\mathbb{E}[\psi(Z;\theta_*)] = 0,$9–$\phi(X;\theta)$0, and often more with tuning, across summarization evaluation, attributed question answering, open-book QA, and side-by-side tests. At $\phi(X;\theta)$1, for open-book QA, chain rule with abstentions achieved mean width $\phi(X;\theta)$2 versus classical $\phi(X;\theta)$3 (ratio $\phi(X;\theta)$4); with binary backoff, chain rule gave $\phi(X;\theta)$5 (ratio $\phi(X;\theta)$6), and power-tuned difference gave $\phi(X;\theta)$7 (ratio $\phi(X;\theta)$8) (Hofer et al., 2024).

SP-CCI reports narrower counterfactual intervals than standard CCI on both a synthetic benchmark and the IHDP semi-synthetic dataset. On IHDP, average prediction interval widths over $\phi(X;\theta)$9 runs were $\mathbb{E}[\psi(Z;\theta)\mid X]$0 for CCI, $\mathbb{E}[\psi(Z;\theta)\mid X]$1 for SP-CCI (LQ), $\mathbb{E}[\psi(Z;\theta)\mid X]$2 for SP-CCI (MQ), and $\mathbb{E}[\psi(Z;\theta)\mid X]$3 for SP-CCI (HQ), with coverage violation rates $\mathbb{E}[\psi(Z;\theta)\mid X]$4, $\mathbb{E}[\psi(Z;\theta)\mid X]$5, and $\mathbb{E}[\psi(Z;\theta)\mid X]$6 for the three SP-CCI variants, respectively (Farzaneh et al., 4 Sep 2025).

Multi-Source Prediction-Powered Inference reports substantial reductions in confidence-region volume. In simulations, MPPI volume was $\mathbb{E}[\psi(Z;\theta)\mid X]$7 versus $\mathbb{E}[\psi(Z;\theta)\mid X]$8 for classic inference in a homogeneous linear setting, $\mathbb{E}[\psi(Z;\theta)\mid X]$9 versus $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$0 in a homogeneous nonlinear setting, $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$1 versus $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$2 under covariate shift with a linear DGP, and $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$3 versus $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$4 under covariate shift with a nonlinear DGP. In the DXA-measured high body fat application, the reported volume reduction was approximately $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$5 for men and approximately $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$6 for women relative to target-only inference (Li et al., 19 Jun 2026).

Wrapper-style GESPI also reports improvements in small-data settings. In the AlphaFold protein structure prediction experiment, OnlyReal abstained on $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$7 of residues with risk $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$8, whereas GESPI achieved an abstention ratio of $$0 = P_U[\phi(X;\theta)] + P_L[\psi(Z;\theta) - \phi(X;\theta)],$$9 with risk approximately $P_U$0, close to the nominal $P_U$1 and guardrail $P_U$2. In AIME25 one-sided testing for LLM win rates, GESPI increased power relative to OnlyReal while Type I error under shuffled answers was at nominal $P_U$3 (Bashari et al., 24 Sep 2025). SynthBH, finally, is reported to achieve empirical FDR close to $P_U$4 and higher power than BH on real data alone in conformal outlier detection benchmarks, while avoiding the severe FDR inflation of naive BH on pooled synthetic $P_U$5-values (Lee et al., 18 Feb 2026).

6. Assumptions, failure modes, and open questions

The central misconception about GESPI is that it requires synthetic labels to be correct. The opposite claim is explicit across the literature: validity is typically guaranteed without requiring the synthetic predictor, autorater, or generator to be unbiased or consistent. What changes with synthetic quality is efficiency, not the formal error target, provided the correction or guardrail mechanism is correctly implemented (Angelopoulos et al., 2023, Song et al., 28 Jan 2026).

The main practical failures arise when the assumptions needed by the correction mechanism are violated. The most prominent is training–inference dependence. “Demystifying Prediction Powered Inference” shows that double-dipping—reusing training data for inference—produces anti-conservative confidence intervals and poor coverage. In the Mosaiks housing study, under double-dipping with $P_U$6, coverage for income and nightlights coefficients fell to approximately $P_U$7, with interval widths less than half of classical intervals (Song et al., 28 Jan 2026).

A second failure mode is distributional mismatch. PPI assumes comparability between labeled and unlabeled samples under MCAR unless MAR-robust variants are used; MPPI requires density-ratio or transport corrections under covariate shift and domain shift; SP-CCI requires overlap and properly specified or boundedly misspecified importance weights; SynthBH requires an extended PRDS-type dependence condition; and GenAI-powered causal inference requires conditional ignorability and positivity in the learned representation space (Li et al., 19 Jun 2026, Farzaneh et al., 4 Sep 2025, Lee et al., 18 Feb 2026, Imai et al., 5 Jul 2025).

A third concern is quality-adaptive overfitting. In Bayesian PPI, tuning the number of regression-tree partitions led to undercoverage for the tree$P_U$8 cases, which were excluded from interval-width summaries (Hofer et al., 2024). In SP-CCI, a poor counterfactual generator can harm calibration, although the debiasing term and RCPS bound are designed to mitigate this (Farzaneh et al., 4 Sep 2025). In multi-task PPI, the theoretical result that affine recalibration is asymptotically equivalent to the original proxy implies that cross-task gains beyond PPI++ are possible only when the shared proxy–truth relationship contains nonlinear structure (Emmenegger et al., 28 May 2026).

The strongest negative result appears in the wrapper-style GESPI paper: Proposition 9 states an impossibility of uniformly safe synthetic usage at level $P_U$9 for arbitrary synthetic distributions, which motivates the use of guardrail slack $P_L$0 and adaptive fallback rather than unconditional pooling (Bashari et al., 24 Sep 2025). This suggests a general principle for the field: synthetic data can be used safely only when the inferential procedure is explicitly designed to degrade gracefully to the real-data baseline.

Open problems are stated repeatedly across the papers. They include MNAR mechanisms, for which both classical complete-case inference and PPI-type estimators are biased (Song et al., 28 Jan 2026); formal conditions for distributional fidelity and calibration when generative models are used as synthetic-data sources (Song et al., 28 Jan 2026, Imai et al., 5 Jul 2025); practical diagnostics for when predictions are “good enough” to beat classical inference in finite samples (Song et al., 28 Jan 2026); and broader finite-sample theory under heterogeneous sources, sequential sampling, or structured outputs (Li et al., 19 Jun 2026, Kilian et al., 23 May 2025). The cumulative literature therefore portrays GESPI as a rigorous but still actively developing paradigm: its basic logic is stable, but its assumptions, diagnostics, and robustness theory remain an important area of ongoing work.