mlsanityData to 14 Oct 2026

General Synthetic-Powered Inference

arXiv 2509.20345 ↗First posted 24 Sep 2025stat.ME cs.LG stat.MLcited by 6

Not on the map. Why?

Abstract (arXiv)

The rapid proliferation of high-quality synthetic data -- generated by advanced AI models or collected as auxiliary data from related tasks -- presents both opportunities and challenges for statistical inference. This paper introduces a GEneral Synthetic-Powered Inference (GESPI) framework that wraps around a broad class of statistical inference procedures to safely enhance sample efficiency by combining synthetic and real data. Our framework leverages high-quality synthetic data to boost statistical power, yet adaptively defaults to the standard method using only real data when synthetic data are of low quality. The error rate of our method remains below a user-specified bound without any distributional assumptions on the synthetic data, and decreases as the quality of the synthetic data improves. This flexibility enables seamless integration with conformal prediction, risk control, hypothesis testing, and multiple testing procedures, all without modifying the base inference method. We demonstrate the benefits of our method on challenging tasks with limited labeled data, including AlphaFold protein structure prediction, and comparing large reasoning models on complex math problems.

Key facts. Cited by 6 papers in the map (data through 14 Oct 2026). Builds on 31 of its 70 references that are in the map. First posted on arXiv 24 Sep 2025.

Builds on

31 of 70 references.

References · 31

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2019A Comprehensive Survey on Transfer Learningcited by 1,160
  3. 2024OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problemscited by 1,136
  4. 2020Classification with Valid and Adaptive Coveragecited by 396
  5. 2022Conformal Risk Controlcited by 299
  6. 2021Distribution-Free, Risk-Controlling Prediction Setscited by 288
  7. 2023Prediction-Powered Inferencecited by 220
  8. 2023Interpreting CLIP's Image Representation via Text-Based Decompositioncited by 196
  9. 2021Testing for Outliers with Conformal p-valuescited by 134
  10. 2023PPI++: Efficient Prediction-Powered Inferencecited by 117
  11. 2018An introduction to domain adaptation and transfer learningcited by 112
  12. 2024AutoEval Done Right: Using Synthetic Data for Model Evaluationcited by 57
19 more
  1. 2015Robust Covariance and Scatter Matrix Estimation under Huber's Contamination Modelcited by 47
  2. 2019Generalized Resilience and Robust Statisticscited by 37
  3. 2025Predictions as Surrogates: Revisiting Surrogate Outcomes in the Age of AIcited by 34
  4. 2024Prediction-Powered Ranking of Large Language Modelscited by 29
  5. 2020Combining Observational and Experimental Datasets Using Shrinkage Estimatorscited by 28
  6. 2021Adaptive Combination of Randomized and Observational Datacited by 22
  7. 2023Transductive conformal inference with adaptive scorescited by 21
  8. 2025Robust Conformal Outlier Detection under Contaminated Reference Datacited by 12
  9. 2025Efficient Randomized Experiments Using Foundation Modelscited by 10
  10. 2021Robust W-GAN-Based Estimation Under Wasserstein Contaminationcited by 9
  11. 2024Reliable Confidence Intervals for Information Retrieval Evaluation Using Generative A.Icited by 9
  12. 2025Synthetic-Powered Predictive Inferencecited by 9
  13. 2024Batch Predictive Inferencecited by 7
  14. 2025Valid Inference with Imperfect Synthetic Datacited by 7
  15. 2023The Real Deal Behind the Artificial Appeal: Inferential Utility of Tabular Synthetic Datacited by 6
  16. 2025GLM Inference with AI-Generated Synthetic Data Using Misspecified Linear Regressioncited by 6
  17. 2015Optimal Multiple Testing Under a Gaussian Prior on the Effect Sizescited by 5
  18. 2023Efficient and Multiply Robust Risk Estimation under General Forms of Dataset Shiftcited by 5
  19. 2025Statistical Methods in Generative AIcited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 1; 2026: 5.

6 papers.

  1. 2026Same Scores, Different Decisions: Evaluating JEV and Language Models for Legal Document Understandingcited by 11
  2. 2025Uncertainty Quantification and Data Efficiency in AI: An Information-Theoretic Perspectivecited by 2
  3. 2026Synthetic-Powered Multiple Testing with FDR Controlcited by 1
  4. 2026Confounding-Valid Conformal Inference for Counterfactual KPIs in Wireless Networkscited by 1
  5. 2026Learning a Size-Weight Frontier for Synthetic-Augmented Inferencecited by 0
  6. 2026Statistical attribute alignment for black-box generative AI via output post-processingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Valid Inference with Synthetic Data via Task Exchangeability
  2. 2025Boosting Statistic Learning with Synthetic Data from Pretrained Large Models
  3. 2025SynthPert: Enhancing LLM Biological Reasoning via Synthetic Reasoning Traces for Cellular Perturbation Prediction
  4. 2026Reasoning-Driven Synthetic Data Generation and Evaluation
  5. 2026Synthesis Without Training: An Inference-Only Pipeline for Tabular, Temporal, and Relational Synthetic Data
  6. 2026Valid Stopping in Adaptive Generator-Verifier Loops
  7. 2026Does Learning Protein Folding Generalize to Broader Reasoning?
  8. 2026Fisher-R1: Training LLM Agents for Reliable Hypothesis Testing