Active Sequential Prediction Mean Estimation

• The paper demonstrates that sequential prediction-powered mean estimation uses predictive surrogates with importance weighting to correct bias from unqueried labels.
• It reveals that a uniform query probability can outperform uncertainty-guided strategies by reducing estimator variance and improving reliability.
• The method provides non-asymptotic confidence bounds and adapts via cross-fitting and control-variate corrections to achieve semiparametric efficiency.

Active sequential prediction-powered mean estimation is a sequential inference setting in which labels are expensive, covariates are observed one at a time, and a predictive model is used whenever the label is not queried. In the formulation revisited in 2026, at each round one must decide the query probability of the ground-truth label upon observing the covariates of a sample; if the label is not queried, the prediction from a machine learning model is used instead (Sfyraki et al., 20 Apr 2026). Closely related work extends prediction-powered inference to growing labeled and unlabeled streams with anytime-valid confidence sequences (Kilian et al., 23 May 2025). By contrast, PPI++ is a computationally lightweight batch methodology rather than a sequential or active data-acquisition method, although its control-variate structure is directly relevant to sequential updating (Angelopoulos et al., 2023).

1. Problem formulation and estimands

In the sequential active formulation, the data arrive as a sequence $(x_t,y_t)$, where $x_t \in X \subset \mathbb{R}$ are covariates drawn from an unknown fixed $\mathbb{P}_X$, and $y_t \in Y \subset \mathbb{R}$ are labels drawn from an unknown conditional distribution $\mathbb{P}_{Y|X}$. The goal is to estimate the mean label

$\mu_y = \mathbb{E}[y_t].$

Because labels are costly, at each round $t$ the algorithm chooses a query probability $\pi_t(x_t)$, draws

$\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$

and observes $y_t$ only if $x_t \in X \subset \mathbb{R}$0. If the label is not queried, the estimator falls back on a predictive model $x_t \in X \subset \mathbb{R}$1, which may evolve over time using past labeled data; the paper assumes $x_t \in X \subset \mathbb{R}$2, where $x_t \in X \subset \mathbb{R}$3 is the filtration generated by the data seen up to time $x_t \in X \subset \mathbb{R}$4 (Sfyraki et al., 20 Apr 2026).

A parallel sequential PPI formulation is asymmetric rather than purely per-round active: there is a large pool of unlabeled examples $x_t \in X \subset \mathbb{R}$5 observed over time, a smaller labeled stream $x_t \in X \subset \mathbb{R}$6, and a black-box predictor $x_t \in X \subset \mathbb{R}$7 that maps $x_t \in X \subset \mathbb{R}$8. The general estimand is

$x_t \in X \subset \mathbb{R}$9

with mean estimation as the main concrete case,

$\mathbb{P}_X$0

For general losses, the estimating equation is based on the subgradient

$\mathbb{P}_X$1

and is decomposed as

$\mathbb{P}_X$2

where

$\mathbb{P}_X$3

For mean estimation, $\mathbb{P}_X$4 does not depend on $\mathbb{P}_X$5 (Kilian et al., 23 May 2025).

This problem class should be distinguished from fixed-time, batch prediction-powered inference. PPI++ assumes that one already has a small labeled sample $\mathbb{P}_X$6, a larger unlabeled sample $\mathbb{P}_X$7 with the same feature distribution, and a black-box predictor $\mathbb{P}_X$8 that outputs $\mathbb{P}_X$9 for any $y_t \in Y \subset \mathbb{R}$0; it does not decide which points to label next (Angelopoulos et al., 2023).

2. Core estimators and control-variate structure

The sequential active estimator analyzed in the 2026 revisit is

$y_t \in Y \subset \mathbb{R}$1

with online recursion

$y_t \in Y \subset \mathbb{R}$2

Its increment

$y_t \in Y \subset \mathbb{R}$3

satisfies $y_t \in Y \subset \mathbb{R}$4, so the update is unbiased in that sense (Sfyraki et al., 20 Apr 2026).

In sequential PPI with growing labeled and unlabeled samples, estimation proceeds by separately estimating the prediction-based term on unlabeled covariates and the rectifier on labeled data. For mean estimation, the paper gives

$y_t \in Y \subset \mathbb{R}$5

and the sequential PPI point estimator

$y_t \in Y \subset \mathbb{R}$6

This is exactly the familiar control-variate estimator, but embedded in a time-uniform inference procedure; $y_t \in Y \subset \mathbb{R}$7 for PPI and $y_t \in Y \subset \mathbb{R}$8 for PPI++-style tuning (Kilian et al., 23 May 2025).

The same control-variate structure appears in batch PPI++. For mean estimation with squared loss $y_t \in Y \subset \mathbb{R}$9, the closed-form estimator is

$\mathbb{P}_{Y|X}$0

Here $\mathbb{P}_{Y|X}$1 recovers standard PPI, $\mathbb{P}_{Y|X}$2 gives the classical labeled-only estimator, and a data-driven $\mathbb{P}_{Y|X}$3 power tunes between the two (Angelopoulos et al., 2023).

These formulations all implement the same basic correction principle: a prediction term is used as a cheap surrogate, and a labeled residual or importance-weighted correction restores validity. A plausible implication is that “prediction-powered mean estimation” is best viewed not as a single algorithm, but as a family of control-variate and rectifier constructions adapted to different sampling regimes.

3. Query-probability design and active acquisition

The central active design problem is the choice of $\mathbb{P}_{Y|X}$4. The revisiting paper studies a mixed policy based on an uncertainty estimate $\mathbb{P}_{Y|X}$5 and a constant query probability: $\mathbb{P}_{Y|X}$6 where

$\mathbb{P}_{Y|X}$7

and $\mathbb{P}_{Y|X}$8 is the mixing parameter. The intended interpretation is direct: $\mathbb{P}_{Y|X}$9 uses only the uncertainty-driven policy, $\mu_y = \mathbb{E}[y_t].$0 ignores uncertainty and queries uniformly at the budget rate $\mu_y = \mathbb{E}[y_t].$1, and intermediate $\mu_y = \mathbb{E}[y_t].$2 interpolates between the two. The operational uncertainty-based rule from prior work is

$\mu_y = \mathbb{E}[y_t].$3

with

$\mu_y = \mathbb{E}[y_t].$4

and clipping to $\mu_y = \mathbb{E}[y_t].$5 (Sfyraki et al., 20 Apr 2026).

The striking empirical observation is that the confidence interval width is smallest when $\mu_y = \mathbb{E}[y_t].$6 is close to $\mu_y = \mathbb{E}[y_t].$7, so the query rule is almost purely constant and the uncertainty-based component is largely removed. In the reported experiments, $\mu_y = \mathbb{E}[y_t].$8 often produced confidence intervals that were slightly narrower than $\mu_y = \mathbb{E}[y_t].$9, and across multiple datasets the uniform policy was comparable to or better than the mixed policy. The paper interprets this as evidence that the uncertainty predictor $t$0 may be brittle or insufficiently reliable in this sequential setting; if uncertainty estimates are noisy or weakly aligned with actual residuals, then mixing them into the query probability can increase estimator variance without providing enough benefit to compensate (Sfyraki et al., 20 Apr 2026).

To formalize this phenomenon, the paper introduces an online convex optimization view. With an oracle $t$1 satisfying

$t$2

it defines the loss

$t$3

and applies Follow-the-Regularized-Leader with regularizer $t$4 over $t$5. The update is

$t$6

with closed form

$t$7

Because $t$8 is minimized by the largest feasible $t$9, the paper’s no-regret interpretation is that the learned query probabilities approach

$\pi_t(x_t)$0

when the policy is chosen obliviously to the current covariates and the objective is only to minimize estimator variance (Sfyraki et al., 20 Apr 2026).

This analysis directly addresses a common misconception: uncertainty-guided querying is not automatically beneficial. In this literature, uncertainty-guided querying is useful only when the uncertainty oracle is very reliable and meaningfully aligned with the conditional residual variance; otherwise, the constant-probability component can dominate.

4. Confidence intervals, confidence sequences, and non-asymptotic guarantees

The revisiting paper’s main technical contribution is a non-asymptotic, data-dependent confidence bound for the sequential estimator. Assuming $\pi_t(x_t)$1 and defining

$\pi_t(x_t)$2

it proves via Freedman’s inequality that for any $\pi_t(x_t)$3, with probability at least $\pi_t(x_t)$4, simultaneously for all $\pi_t(x_t)$5,

$\pi_t(x_t)$6

The bound contains a burn-in term $\pi_t(x_t)$7, a variance-adaptive term controlled by $\pi_t(x_t)$8, and a clipping term depending on $\pi_t(x_t)$9 and $\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$0. When $\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$1 dominates the boundedness term, the rate behaves as

$\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$2

and using the trivial bound $\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$3 gives a worst-case $\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$4 rate (Sfyraki et al., 20 Apr 2026).

Sequential PPI work addresses a different inferential question: validity uniformly over time. The anytime-valid, Bayes-assisted extension uses Ville’s inequality and the method of mixtures to construct confidence sequences. The PPI confidence set is obtained by building a confidence sequence for $\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$5 and inverting: $\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$6 For mean estimation, the resulting interval is

$\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$7

The guarantee is

$\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$8

so one can peek continuously without losing validity (Kilian et al., 23 May 2025).

Fixed-time batch inference remains important as a baseline. PPI++ replaces the original PPI approach of testing $\xi_t \sim \mathrm{Bernoulli}(\pi_t(x_t)),$9 over a grid of candidate values by computing a single point estimate, estimating its asymptotic covariance, and placing standard normal error bars around it. For scalar mean estimation, the interval reduces to

$y_t$0

with asymptotic coverage

$y_t$1

coordinatewise (Angelopoulos et al., 2023).

Taken together, these results separate three inferential regimes: non-asymptotic fixed-horizon bounds for active sequential querying, anytime-valid confidence sequences for growing streams, and asymptotic fixed-time intervals for batch semi-supervised inference.

5. Statistical efficiency, calibration, and learned predictors

A foundational theoretical development reframes PPI as an M-estimation or Z-estimation problem. The target parameter $y_t$2 is defined by

$y_t$3

and PPI replaces the unavailable responses for most units by a predictor $y_t$4 and then bias-corrects using labeled residuals. For mean estimation, where

$y_t$5

the PPI estimator becomes

$y_t$6

Under simple random sampling without replacement, the score is design-unbiased for the infeasible full-data estimating equation (Lee et al., 7 Jun 2026).

The mean-specialized asymptotic variance is

$y_t$7

equivalently

$y_t$8

This decomposition makes explicit that efficiency depends on both the variability of the regression function and the residual variability left after prediction (Lee et al., 7 Jun 2026).

The same paper identifies the efficient influence function in the semiparametric model with observed data $y_t$9 and known labeling fraction $x_t \in X \subset \mathbb{R}$00. For mean estimation, the efficient influence function is

$x_t \in X \subset \mathbb{R}$01

and the efficiency bound is

$x_t \in X \subset \mathbb{R}$02

PPI attains this bound when the predictor is score-calibrated, which for the mean means

$x_t \in X \subset \mathbb{R}$03

This point is conceptually important: misspecification affects efficiency, not consistency, whereas calibration governs attainment of the semiparametric lower bound (Lee et al., 7 Jun 2026).

When the predictor is learned from the labeled data, inferential correction becomes essential. The paper proves that vanilla PPI with learned $x_t \in X \subset \mathbb{R}$04 can have severe undercoverage when $x_t \in X \subset \mathbb{R}$05 is small, while cross-fitting and a single-fit variance-corrected variant largely restore nominal coverage and still give shorter intervals than classical estimation. For mean estimation, the cross-fitted estimator is

$x_t \in X \subset \mathbb{R}$06

and asymptotic normality requires only $x_t \in X \subset \mathbb{R}$07-consistency of the out-of-fold predictor; no Donsker, entropy, or stability rate like $x_t \in X \subset \mathbb{R}$08 is required for the mean (Lee et al., 7 Jun 2026).

Batch PPI++ addresses a related but distinct adaptation problem through “power tuning.” For mean estimation, the optimal asymptotic coefficient is

$x_t \in X \subset \mathbb{R}$09

and the plug-in $x_t \in X \subset \mathbb{R}$10 is near $x_t \in X \subset \mathbb{R}$11 when $x_t \in X \subset \mathbb{R}$12 is highly predictive and near $x_t \in X \subset \mathbb{R}$13 when $x_t \in X \subset \mathbb{R}$14 is poor or uncorrelated. This yields the paper’s central message that the method automatically adapts to prediction quality and is typically no worse than classical inference (Angelopoulos et al., 2023).

6. Related formulations, scope boundaries, and recurring misconceptions

The literature grouped under this topic includes several adjacent formulations that are not identical. PPI++ is batch inference rather than a sequential or active data-acquisition method; it gives a principled way to combine a growing labeled set with cheap predictions, and its tuning parameter can be updated as more labels arrive, but the paper itself does not provide an active query policy or sequential stopping rule (Angelopoulos et al., 2023).

A separate finite-population line formulates the problem as active measurement. There is a finite population $x_t \in X \subset \mathbb{R}$15 with $x_t \in X \subset \mathbb{R}$16, each unit $x_t \in X \subset \mathbb{R}$17 has an unknown nonnegative quantity $x_t \in X \subset \mathbb{R}$18, and the goal is to estimate the population total

$x_t \in X \subset \mathbb{R}$19

with the population mean obtained by dividing by $x_t \in X \subset \mathbb{R}$20. At time $x_t \in X \subset \mathbb{R}$21, an AI model produces predictions $x_t \in X \subset \mathbb{R}$22 for unlabeled units and defines a proposal distribution $x_t \in X \subset \mathbb{R}$23 over $x_t \in X \subset \mathbb{R}$24. The core estimator is

$x_t \in X \subset \mathbb{R}$25

which is unbiased for any proposal supported on the unlabeled set. The paper then aggregates roundwise estimators as

$x_t \in X \subset \mathbb{R}$26

with the key covariance identity

$x_t \in X \subset \mathbb{R}$27

and constructs confidence intervals by a martingale central limit theorem using

$x_t \in X \subset \mathbb{R}$28

This is sequential and active, but it operates in a finite-population, without-replacement Monte Carlo setting rather than the i.i.d. online query-probability model (Hamilton et al., 2 Jul 2025).

Another neighboring framework is collaborative online personalized mean estimation. There are $x_t \in X \subset \mathbb{R}$29 agents, each with a private distribution $x_t \in X \subset \mathbb{R}$30 over $x_t \in X \subset \mathbb{R}$31 and mean

$x_t \in X \subset \mathbb{R}$32

and each agent can actively query another agent for its current empirical average. The ColME algorithm identifies, on the fly, which agents are likely in the same mean class and aggregates only those agents’ estimates. Under Restricted-Round-Robin, the mean-estimation time is

$x_t \in X \subset \mathbb{R}$33

and the collaborative method can be up to $x_t \in X \subset \mathbb{R}$34 times faster than local estimation after class discovery. This is framed as collaborative personalized estimation rather than prediction-powered inference, but the paper explicitly notes that it is very close in spirit to active sequential prediction-powered estimation because queried auxiliary summaries can “power” the final estimator once they are deemed trustworthy (Asadi et al., 2022).

Several misconceptions recur across these papers. One is that sequential prediction-powered inference must be uncertainty-driven; the revisiting analysis instead shows that under both theory and experiments, the best-performing policies are often the ones that behave nearly like uniform sampling at the budget rate (Sfyraki et al., 20 Apr 2026). Another is that using predictions directly is sufficient for valid inference; across the literature, validity is restored by rectification, importance weighting, cross-fitting, variance correction, or martingale-based confidence constructions rather than by raw prediction alone (Kilian et al., 23 May 2025). A further misconception is that all “active” prediction-powered methods solve the same problem. The papers instead address distinct regimes: online query-probability control for a stream of covariates and costly labels, anytime-valid semi-supervised inference with growing labeled and unlabeled data, finite-population active measurement without replacement, and collaborative querying across multiple sample streams (Sfyraki et al., 20 Apr 2026).

These distinctions suggest a unifying view. Active sequential prediction-powered mean estimation is not a single estimator but a family of methods that combine predictions with selective labeling to estimate a mean under explicit budget constraints, while differing in sampling design, validity notion, and efficiency analysis.