acceptodds
Under review as a conference paper at ICLR 2027

Transfer Calibrated Prediction Powered Inference

Abstract

Prediction-powered inference (PPI) and its power-tuned extension (PPI++) improve confidence intervals by combining a small gold-standard labeled sample with a large AI model's predictions. Its efficiency gain relies on low residual variance, which may not hold if the predictor is pre-trained on a different source domain. We propose Transfer Calibrated Prediction-Powered Inference (TC-PPI), adapting the source-domain predictor to the target domain using gold-standard samples through cross-fitting. This approach supports various adaptation methods, such as sparse linear calibration, LoRA, and fine-tuning. Our jointly tuned cross-fit estimator, Joint-TC-Cross-PPI++, maintains unbiasedness and is simultaneously at least as efficient as classical inference, PPI, and PPI++, thereby protecting against negative transfer. We provide high-dimensional MSE bounds for calibration and show empirical improvements over baseline methods across various real-world applications.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Prediction Powered Inference, Transfer Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MbELOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adapts a source predictor to a target population using a correction learned from labeled target data. Cross-fitting allows all labeled observations to contribute to rectification. A joint estimator separately weights the source prediction and the learned correction. The analysis gives fixed-weight unbiasedness, asymptotic distributions, an asymptotic no-harm comparison, and sparse-calibration risk bounds. Experiments evaluate mean and regression targets across several data modalities.

Critical strengths

Separately weighting the source prediction and transfer correction provides a useful safeguard. The oracle comparison follows transparently from the nested estimator classes, and the sparse-calibration bounds quantify the cost of learning the correction.

The empirical study covers several inference targets and modalities and reports coverage and small-sample failures. For example, the Bikeshare results show a relative-MSE reduction from 0.211 for PPI++ to 0.139 for the joint estimator at n=100.

Critical weaknesses

The abstract overstates the guarantees for the feasible jointly tuned estimator. Proposition 3 establishes unbiasedness for fixed weights, while the implementation estimates weights using the same gold responses used for rectification. Cross-fitting the correction does not automatically preserve exact unbiasedness after this additional estimation step. Theorem 3 gives an asymptotic comparison, not finite-sample protection against negative transfer. The body is more careful, but the abstract needs correction.

The comparison with existing Cross-PPI is inaccurate. The paper contrasts it with pretrained-model adaptation, but Zrnic and Candès explicitly allow fine-tuning an off-the-shelf model in Section 3.1. Appendix F.5 evaluates a scratch-trained implementation, which does not resolve this overlap. The contribution should be positioned through joint tuning and the additional risk analysis.

Questions for authors

The abstract attributes exact unbiasedness to the jointly tuned estimator, but Proposition 3 assumes fixed weights. What argument justifies exact unbiasedness when the weights are estimated from the same gold responses used for rectification? If no such argument applies, please qualify the abstract accordingly.

How often did the default [0,1]^2 weight constraints bind in the experiments, and do the main efficiency comparisons change when the weights are unconstrained? This would clarify the empirical relevance of Theorem 3's unrestricted oracle comparison.

3 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for taking the time to review our paper.

[W1/Q1] The abstract overstates the guarantees for the feasible jointly tuned estimator. The body is more careful.

We will change the wording in the abstract in the final revision, as this is a minor stylistic concern. As you have noted, we make the distinction clear in the body of the paper. The method is unbiased with oracle fixed λ\lambda, while it still remains consistent with estimated λ\lambda; both properties are exactly the same as in PPI++.

[W2] The comparison with existing Cross-PPI is inaccurate as the comparison uses scratch-trained implementation. The contribution should be positioned through joint tuning and the additional risk analysis.

We have used Cross-PPI as described by the authors of [1]. The paper's theory treats the training algorithm as a black box, requiring only that the fitted models be stable. A fine-tuned model, or a trained-from-scratch model, is just the learned predictor, and there are no guarantees relative to the pretrained model and no consideration of the costs or benefits of adaptation. The cross-PPI paper focuses on training an algorithm using the gold sample data. The methods, theory, applications, and code in [1] all point to that, and we have used the method exactly as the authors intended.

In Figure 8 (Appendix F.5), Cross-PPI and TC-Cross-PPI are compared apples-to-apples since they both start from the source model, use the same gold labels and cross-fitting folds, and neither is power-tuned. Cross-PPI trains an instance of the source model from scratch on the gold labels, while TC-Cross-PPI keeps the source model fixed and learns only a sparse correction from the same gold labels. In this experiment, we sought to isolate the value of starting from the source model. Joint TC-Cross-PPI++, also shown in Figure 8, then keeps the source prediction and the learned correction as separate terms and power-tunes them jointly. We will make this point clearer in a revised version.

[Q2] How often did the default [0,1]^2 weight constraints bind in the experiments, and how would unconstraining them perform

A constraint on the weights for general M-estimation targets guarantees convexity (the Bradley-Terry applications). As in the original PPI++ paper [2]: weights outside [0,1][0,1] can make the rectified objective nonconvex. For mean targets, unconstrained weights are valid, and removing the box matters only when the source-to-target transfer depends little on the covariates (e.g., an affine relationship between the prediction scores, which rescaling captures). The table below reports mean Classical MSE / mean Method MSE with the [0,1]2[0,1]^2 box →\to without it. Census Income and Capital Bikeshare change the most, while all other settings barely change.

ApplicationnnPPI++Joint TC-Cross-PPI++
Census Income2001.63 →\to 1.811.72 →\to 1.72
Capital Bikeshare1005.38 →\to 10.819.20 →\to 10.44
Camelyon17 (WILDS)1001.18 →\to 1.181.39 →\to 1.39
PovertyMap (WILDS)2003.23 →\to 3.233.79 →\to 3.78
ProteinGym2001.34 →\to 1.341.39 →\to 1.39
AutoEval-ProteinGym5001.44 →\to 1.441.65 →\to 1.65
AfriSpeech-200 (Yoruba)5001.06 →\to 1.091.33 →\to 1.30
ProteinGym-ESM: ESM-2 8M5001.18 →\to 1.181.47 →\to 1.47
ProteinGym-ESM: ESM-2 35M5001.25 →\to 1.251.56 →\to 1.56
ProteinGym-ESM: ESM-2 150M5001.57 →\to 1.571.94 →\to 1.94

For the Bradley-Terry targets, following [2], we kept each weight in [0,1][0,1], but the region can be loosened to 0≤λs+λΔ≤10 \le \lambda_s + \lambda_\Delta \le 1. PPI++ restricts its single weight to [0,1][0,1] so that the rectified GLM loss stays convex [2]; with our two pseudo-label terms in Eq. (61), the same argument depends only on the total weight λs+λΔ\lambda_s + \lambda_\Delta, allowing individual weights to be outside of [0,1][0,1]. The loss then stays convex, so its minimizer is the root mentioned in Theorem 10. Below we report the percentage of trials in which each constraint type (box vs sum) binds for Joint TC-Cross-PPI++, and report the mean Classical MSE / mean Method MSE for each.

Applicationnnλ^s\hat\lambda_s at bound (box)λ^Δ\hat\lambda_\Delta at bound (box)Sum at bound (sum constraint)box →\to sum constraint
Chatbot Arena5008%44%6%1.04 →\to 1.04
Chatbot Arena10000%36%0%1.08 →\to 1.08
Chatbot Arena20000%18%0%1.11 →\to 1.12
AutoEval MT-Bench1001%45%1%1.15 →\to 1.16
AutoEval MT-Bench2000%36%0%1.08 →\to 1.08
AutoEval MT-Bench5000%10%0%1.07 →\to 1.07

Under the box constraint, binding mostly occurs through λ^Δ\hat\lambda_\Delta, while the sum constraint rarely binds. The metric is essentially unchanged.


References

[1] Zrnic & Candès. Cross-prediction-powered inference. PNAS, 2024.

[2] Angelopoulos, Duchi & Zrnic. PPI++: Efficient prediction-powered inference. 2023.

Thank you for the additional experiments. My main concern about the Cross-PPI comparison remains as it explicitly allows fine-tuning a pretrained model. I maintain my original rating.

AuthorsOpenReview ↗

Concerns about AI Usage

We'd like to note that this reviewer's original review was rated 100% AI by Pangram (https://www.pangram.com/history/167b3d3c-5843-4d34-87d0-54e33331b842). As this reviewer has continued to provide nothing substantive or constructive and, based on the latest response, seems not to comprehend our original response in any meaningful sense (we already addressed the concern in our initial response), we have nothing further to add to this reviewer thread.

Reviewer_jYw9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies prediction-powered inference when the available predictor was trained on a different source domain. It uses labeled target data to learn a correction to the source predictor, with cross-fitting so that all labeled examples can still be used for inference. A joint version separately weights the source prediction and the learned correction. Experiments cover several application domains.

Critical strengths
  1. The cross-fitting design uses all labeled target examples for rectification while ensuring that each example is evaluated with a correction learned without that example.

  2. The experiments cover tabular data, speech, medical imaging, satellite imagery, protein models, and LLM evaluation. The paper reports MSE, interval width, coverage, and several failure cases.

Critical weaknesses
  1. The baseline comparison is incomplete. ProteinGym and Arena provide multiple predictors, making Multi-source PPI++ a natural baseline, but it is not reported even though the supplementary code computes it. Cross-PPI is evaluated only on two tabular datasets. The current experiments therefore do not establish the proposed method’s advantage over its closest alternatives.

  2. The uncertainty of the empirical comparisons is not reported. The main figures and Table 1 provide no error bars or Monte Carlo standard errors, although some differences are very small. For example, the relative MSE on AutoEval MT-Bench is 0.933 for Joint TC-Cross-PPI++ and 0.932 for PPI++. It is unclear which small differences are reproducible and which are caused by repeated subsampling noise.

Questions for authors

none

3 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for taking the time to review our paper.

[W1] Need to compare against Multi-Source PPI++. Why is Cross-PPI evaluated only on two tabular datasets.

Below, we compare Multi-Source PPI++ [1] against a multi-source version of our method. Our "Multi-Source Joint TC-Cross-PPI++" gives each source predictor its own cross-fit correction and tuned joint weights, (λs,λΔ)(\lambda_s, \lambda_{\Delta}), so for mm source predictors, we have 2m2m weights. If all the λΔ\lambda_{\Delta}'s are set to zero, then we recover Multi-Source PPI++. The table reports the mean Classical MSE / mean Method MSE, with the higher value being bolded and italicized others in the same row that are statistically equivalent according to a 95% paired bootstrap. In brackets, we also report the median Classical MSE / median Method MSE. Most results align whether we use the mean or median, but the latter helps us better understand the impact of outliers; see the Chatbot Arena experiment at n=500n = 500: with only 500 labeled comparisons, some models appear in just a few, which artificially inflates the mean-based ratio.

Application (sources)nnMulti-source PPI++Multi-source Joint TC-Cross-PPI++
ProteinGym-ESM (ESM-2 8M, 35M, 150M)5001.56 (0.96) [1.33]1.90 (0.97) [1.66]
ProteinGym-ESM (ESM-2 8M, 35M, 150M)10001.78 (0.98) [1.55]2.14 (0.99) [1.96]
ProteinGym-ESM (ESM-2 8M, 35M, 150M)20001.55 (1.00) [1.50]1.84 (1.00) [1.92]
AutoEval-ProteinGym (VESPA, VESPAl, VespaG)5001.40 (0.90) [1.30]1.61 (0.89) [1.56]
AutoEval-ProteinGym (VESPA, VESPAl, VespaG)10001.43 (0.90) [1.54]1.58 (0.89) [1.69]
AutoEval-ProteinGym (VESPA, VESPAl, VespaG)20001.44 (0.90) [1.61]1.59 (0.89) [2.05]
AutoEval-ProteinGym, 8 sources (7 candidate families + VESPA)5001.81 (0.89) [1.94]1.85 (0.87) [1.81]
AutoEval-ProteinGym, 8 sources (7 candidate families + VESPA)10001.84 (0.90) [1.81]1.93 (0.89) [2.03]
AutoEval-ProteinGym, 8 sources (7 candidate families + VESPA)20001.79 (0.90) [2.04]1.95 (0.90) [2.23]
Chatbot Arena, Bradley-Terry (17 LLM judges)5002.84 (0.95) [1.32]2.80 (0.95) [1.27]
Chatbot Arena, Bradley-Terry (17 LLM judges)10001.18 (0.96) [1.19]1.17 (0.96) [1.19]
Chatbot Arena, Bradley-Terry (17 LLM judges)20001.15 (0.97) [1.21]1.17 (0.96) [1.21]

Regarding Cross-PPI [2]: Cross-PPI learns the predictor on the gold labels with cross-fitting. We evaluate it on Census and Bikeshare, where the source model class (gradient-boosted trees) can be trained on n≤2000n \le 2000 labels (see Appendix F.5). For the neural sources (DenseNet, ResNet, wav2vec2, ESM-2, LLM judges), training a comparable model on a few hundred labels is not meaningful, and adapting the pretrained source is exactly the setting our method targets.

[W2] Missing error bars in empirical comparisons

In the final revision, we will include error bars (95% paired bootstrap intervals) in all of the figures and report standard errors in Table 1.

AuthorsOpenReview ↗

Following-up with Reviewer

We thank Reviewer jYw9 for taking the time to review our work and provide constructive feedback. As the Author-Reviewer discussion period draws to a close, we would appreciate the opportunity to address any remaining questions or concerns.

Thank you for providing the additional experiments and the clarification regarding Cross-PPI. These largely address my concerns about the baseline. Please ensure that error bars and standard errors are included in the revised manuscript. I will maintain my rating of 3 (weak acceptance).

Reviewer_hkNMOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is interesting, and is a natural generalization of existing work on PPI with transfer. The main contribution as I understand it is to extend to a joint optimization of the linear correction factor over the source prediction and the rectifier, and also to extend to higher dimensions using sparsity. The authors present theoretical results concerning the estimators which are interesting, and empirical results on several datasets.

Critical strengths
  • The estimators that the authors present are interesting and potentially valuable, if their relationship to existing methods can be further clarified and compared.
  • I think the theoretical analysis is particularly worthwhile. I think all of the minimax optimality, sparsity, and non-asymptotic results are interesting.
Critical weaknesses

While the paper has interest propositions listed above, it would be helpful if they can clarify the difference with existing work in PPI and transfer, most notably RePPI, which I was surprised to see is omitted.

It seems to me that the main advancement over RePPI is joint optimization of the hyperparameter, and perhaps consideration of sparsity when fitting this parameter. Is this correct? Please clarify if not. I would recommend:

  • An empirical comparison to RePPI
  • A theoretical discussion of when this method is better/worse than what is presented in RePPI. For instance, in which regime is the joint optimization beneficial?

If this is done, I am happy to revisit my score.

Questions for authors
  • It seems that TC-Cross PPI and TC-Cross PPI++ match Joint TC-Cross PPI++ most of the time, which seems interesting. Do you have any idea why? Do you think most of the benefit comes from those?

  • In terms of the writing, it wasn't clear on my first read what your main contribution was and how it differs from existing work. I would recommend explaining your "main contributions" in the introduction rather than "core methodology"? Also stylistically I would prefer a "literature review" rather than explaining your "key differentiators", but this is of course a nit.

  • It would also be helpful if you can clarify the relationship between the results in section 4 and those presented in https://arxiv.org/abs/2603.27414 since the stated results are similar in spirit. For instance, how do these results go beyond those? I just think it would be helpful to clarify what the improvements are.

  • What are the implications of your non-asymptotic results for when people should use this method compared to others?

5 comments
AuthorsOpenReview ↗

Response to Reviewer [1/3]

We thank the reviewer for taking the time to review our paper.

[W1] Compare against RePPI (a) theoretical discussion and (b) empirical comparison.

Theoretical Comparison: We thank the reviewer for bringing the RePPI paper [1] to our attention, which was an omission on our part. Indeed, RePPI and TC-PPI are similar in that they both use labeled gold data to improve the pretrained predictions used for inference through careful use of cross-fitting. However, TC-PPI approaches the problem differently, creating a different methodology and theoretical analysis.

While conceptually related, the different choices we make lead to better empirical performance of TC-PPI in a wide range of datasets, as described below (we also compare with another highly related recent proposal called Calibeating PPI[2]). RePPI aims to learn the conditional expectation of the score function, such as gradient of a loss or log-likelihood, given XX and pre-trained predictions Y^\hat{Y} from the labeled data. Since the score generally depends on the unknown target parameter θ\theta, computing this requires a pilot estimate of the target parameter θ^\hat{\theta} from part of the gold samples, while the actual recalibration happens in another part of the gold samples. In contrast, we take a transfer-learning view and learn a low-complexity discrepancy function Δ∗\Delta^* from the target labeled data to calibrate our predictions Y^\hat{Y} directly. This makes the transferred outcome model independent of the target parameter we wish to infer and thus saves expensive gold samples. In a cross-fitting framework, both approaches use all nn gold labels for the PPI rectification stage, but our TCPPI can use more target data ((K−1)n/K(K-1)n/K for a K-fold cross-fitting) for learning the transfer correction, while RePPI needs to spend a portion of the data to learn pilot θ^\hat{\theta}, which cannot be reused for learning the score function.

Further, there is a difference in terms of information used for adaptation in the target domain. In the notation of RePPI, the high-dimensional information-rich covariates WW (e.g., text and image embeddings) are used only through source predictions Y^\hat{Y} and not in the recalibration step. In the main RePPI algorithm, the recalibration step only uses covariates that are needed in the estimating equation for the parameter of interest. However, in TCPPI, our core assumption is that due to domain shift or having a different context, the pre-trained source model predictions Y^\hat{Y} systematically differ from the response in the target domain by a discrepancy function that can be learned using XX and WW. In TCPPI, we are able to use these covariates WW to learn low-complexity adaptation of the source predictions. Our theoretical results also differ. The RePPI estimator is shown to be asymptotically efficient provided the conditional score is accurately learnt from the target labels in [1]. Our calibration through Joint TC-Cross-PPI++ does not aim for asymptotic efficiency but instead aims to improve efficiency simultaneously over classical and source-only PPI++. Our theory further quantifies the finite-sample cost of learning the source-to-target correction and, under the assumption of sparse linear discrepancy, gives non-asymptotic rates and a matching minimax lower bound.

Empirical Comparison: Below we empirically compare Joint TC-Cross-PPI++ against RePPI [1] and the best method from [2] (i.e., one among LinearCal, IsoCal, and AutoCal). The table reports the mean Classical MSE / mean Method MSE, where a higher value means the method's MSE is smaller. The higher value is bolded and italicized for others in the same row that are statistically equivalent according to a 95% paired bootstrap. We observe that in almost every application, Joint TC-Cross-PPI++ outperforms [1] and [2]. In Census Income and Capital Bikeshare, the source predictor already captures most of the signal, leaving little room for transfer calibration to help.

ApplicationnnRePPIBest Calibeating methodBest CalibeatingJoint TC-Cross-PPI++
Census Income2001.81 (0.93)AutoCal1.83 (0.93)1.72 (0.92)
Capital Bikeshare10010.39 (0.99)LinearCal11.16 (0.96)10.44 (0.95)
Camelyon171001.15 (0.97)LinearCal1.19 (0.97)1.39 (0.94)
PovertyMap2003.23 (0.94)AutoCal3.23 (0.96)3.78 (0.94)
ProteinGym2001.30 (0.95)LinearCal1.33 (0.95)1.39 (0.96)
AutoEval-ProteinGym5001.40 (0.90)LinearCal1.44 (0.90)1.65 (0.91)
AfriSpeech-200 (Yoruba)5001.07 (0.94)AutoCal1.10 (0.95)1.30 (0.93)
ProteinGym-ESM: ESM-2 8M5001.18 (0.97)IsoCal1.19 (0.97)1.47 (0.98)
ProteinGym-ESM: ESM-2 35M5001.24 (0.96)AutoCal1.26 (0.96)1.56 (0.97)
ProteinGym-ESM: ESM-2 150M5001.55 (0.96)AutoCal1.61 (0.96)1.94 (0.98)

We will add these results to the revised paper.

AuthorsOpenReview ↗

Response to Reviewer [2/3]

[Q1] Any thoughts on why TC-Cross-PPI(++) matches Joint TC-Cross-PPI++ most of the time

This is because joint TC-Cross-PPI++ optimizes the power-tuning parameter vector λ\lambda on the source predictor and transfer correction in order to minimize the variance. Therefore, the method searches for the best λ\lambda in a class of unbiased estimators that includes the classical estimator, source-only PPI++, and TC-Cross-PPI++ as special cases. As we note in the main paper: \begin{align*} (0,0)^\top: \text{classical inference}, \quad (1,0)^\top&: \text{Source PPI using }f_s, \quad (1,1)^\top: \text{TC--Cross-PPI},\ (\lambda,0)^\top: \text{Source PPI++ using }f_s, & \quad (\lambda,\lambda)^\top: \text{TC--Cross-PPI++}. \end{align*} Therefore, the performance of Joint-TC-Cross-PPI++ is expected to match the performance of TC-Cross-PPI++ if the transfer calibration works well. The joint power tuning only provides protection against harmful negative transfer by having the ability to fall back to source-only PPI++ if the transfer is ineffective, or to classical inference if PPI is entirely ineffective. The empirical results show that TC-Cross-PPI++ ends up being the best often, and the Joint TC-Cross-PPI++ provides little improvement over it. This observation suggests that transfer calibration is successful in most of our applications.

[Q2] In the introduction discuss "main contributions" rather than "core methodology". Stylistically, "literature review" is better than explaining your "key differentiators"

Thank you for the suggestion. We agree that we should discuss main contributions more clearly in the introduction. We will address all presentation and stylistic issues in the final revision.

[Q3] It would also be helpful if you can clarify the relationship between the results in Section 4 and those presented in [3] since the stated results are similar in spirit. For instance, how do these results go beyond those?

MultiPPI [3] and our Section 4 address different sources of finite-sample cost. Indeed, the theoretical results in Section 4 share the spirit of those in MultiPPI. Both decompose the MSE of the proposed estimator as the sum of oracle risk and excess risk. TC-PPI controls the local debiasing rate, while MultiPPI controls the covariance estimation error rate ∥Σ−Σ^∥op\|\Sigma-\widehat\Sigma\|_{\rm op}. However, we clarify that the two models are designed to solve completely different problems.

First, MultiPPI aims to estimate E(a⊤X)\mathbb{E}(a^\top X) and allocate queries (i.e., sample sizes nIn_I and weights λI\lambda_I) under budget BB when the data variance is fixed, while TC-PPI works on transfer learning for a high-dimensional regression and estimates a sparse correction vector δ∗\delta^* and power vector λ\lambda. The latter allows a domain shift when the source dataset offers a powerful but biased fit fsf_s. MultiPPI may use an external sample to estimate parameters, but still under a common distribution for XX.

In terms of minimax results, MultiPPI proves the minimax optimal bound only when Σ\Sigma is known, and there is no matching bound when Σ^≠Σ\widehat\Sigma\ne\Sigma. TC-PPI proves a matching bound up to universal factors and logarithmic factors in s,ps,p for unknown transfer correction coefficient δ∗\delta^*.

AuthorsOpenReview ↗

Response to Reviewer [3/3]

[Q4] What are the implications of your non-asymptotic results for when people should use this method compared to others?

Recall that Theorem 4 proves the following nonasymptotic upper bound

\widehat\mu_{\rm TC\text{-}Cross,\lambda_0} \mid\mathcal D_s \right) \le \frac{\sigma_\epsilon^2}{n} + \frac{V_t}{n+N} + K\frac{N}{n(n+N)}r_n,$$ The first two terms are the "oracle" PPI++ variance that requires the knowledge of target domain function $f_t$. The third term quantifies the finite sample cost of the transfer-calibration. How high this cost is depends on $r_n$, which measures the MSE of estimating the transfer correction. If this error is large, then the bound can be much higher than oracle PPI++ variance (which we emphasize is different from the source-only PPI++ variance in our settings). In the special case of sparse linear calibration with lasso, $r_n\lesssim\frac{\sigma_\epsilon^2s\log p}{n_{\rm tr}}$. Therefore, if $r_n$ is small, then the error due to transfer correction is small, and TC-Cross-PPI++ attains the oracle PPI++ variance. In comparison, the optimally tuned ordinary source PPI++ has variance $$\operatorname{MSE}\left(\widehat\mu_{\mathrm{source\text{-}PPI++}}\right)=\frac{\sigma_\epsilon^2+V_t}{n}-\frac{N}{n(n+N)}\frac{\operatorname{Cov}_t\{Y,f_s(X)\}^2}{\operatorname{Var}_t\{f_s(X)\}}.$$ Using the notation $A_s=\frac{\operatorname{Cov}_t\{Y,f_s(X)\}^2}{\operatorname{Var}_t\{f_s(X)\}}$ for convenience, we have the source PPI++ MSE as $\operatorname{MSE}(\widehat{\mu}_{\text{source-PPI++}}) = \frac{\sigma_\epsilon^2+V_t}{n} - \frac{N}{n(n+N)}A_s.$ We can also rewrite our bound on the TC-Cross-PPI++ for ease of comparison as $\operatorname{MSE}\left(\widehat\mu_{\mathrm{TC\mathrm{-}Cross},\lambda_0}\mid\mathcal D_s\right)\le\frac{\sigma_\epsilon^2+V_t}{n}-\frac{N}{n(n+N)}\left(V_t-Kr_n\right).$ Now we can directly compare that TC-Cross-PPI++ has a lower variance than source PPI++ if $$Kr_n<V_t-A_s.$$ This provides a direct comparison of when TC-Cross-PPI++ performs better than source-only PPI++. When we expect the cost of transfer adaptation to be low, $Kr_n$ will be small, and then TC-Cross-PPI++ is beneficial. Our joint power tuning further provides downside risk protection in case the transfer is harmful and is simultaneously at least as efficient as classical, source PPI++ and TC-cross-PPI++. We recommend users always use the joint-TC-cross-PPI++ as there is asymptotically no downside (except for finite sample error of estimating $\lambda$), and there can often be strong gains as our empirical results show. ---- **References** [1] Ji, Lei & Zrnic. Predictions as Surrogates: Revisiting Surrogate Outcomes in the Age of AI. 2025. [2] van der Laan & van der Laan. Calibeating Prediction-Powered Inference. 2026. [3] Cowen-Breen, Agarwal, Bates, Cohen, Eisenstein, Globerson & Fisch. Multiple-Prediction-Powered Inference. ICLR 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.