Research map / Probabilistic and scientific modeling / Molecules and proteins
Prediction-Powered E-Values
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Extends prediction-powered inference from Z-estimation to e-values, giving a prediction-powered e-value construction that yields valid hypothesis tests, confidence sequences and e-value-based algorithms under costly labels.
Problem
Prediction-powered inference is largely restricted to Z-estimation problems (means, quantiles, regression coefficients), leaving many e-value-based procedures (anytime-valid testing, change-point detection, causal discovery) unsupported when expensive labels are scarce.
Approach
Debias the product of predictable e-value components using randomly sampled expensive labels with known collection probabilities (active-inference style), so the prediction-powered components keep the same conditional expectation as the original e-value components; any e-value-based procedure can then be made prediction-powered by substituting these components.
Key results
- Prediction-powered e-values give tighter valid confidence sequences than labelled-only and vanilla PPI on diabetes prevalence, while imputation fails to cover the true mean.
- In forest cover monitoring under poisoning, prediction-powered methods detect drift much faster than the labelled-only baseline with the same labels; imputation raises false alarms.
- In causal discovery with 10% labelling budget, the method detects over half of the edges, matching the full-data oracle, whereas labelled-only detects none.
Surprising findings
- Imputation detects distribution shift earlier but does so before the null is actually violated, producing high-confidence false alarms.
Limitations
- Requires components of e-values to be predictably bounded, forcing bet truncation and a loss of power.
- Measure-of-confidence-set bounds are only given under heavy assumptions and are not shown to be nonvacuous.
- Causal discovery relies on asymptotic p-values with batching, which are not truly sequential or nonasymptotic.
Experimental setup
Datasets: diabetes survey data (dataset of cited work) forest cover type dataset synthetic DAG data with 6 variables
Benchmarks: diabetes prevalence estimation forest cover online risk monitoring change-point detection on forest cover stream causal discovery on synthetic DAG
Metrics: confidence interval width e-value growth coverage detected edges
Baselines: only labelled samples imputation vanilla prediction-powered inference Fisher's z partial correlation test (PC algorithm)
Builds on: prediction-powered inference active statistical inference e-values aGRAPA PToE calibrator
Closest papers by meaning
- Prediction-Powered Adaptive Shrinkage Estimation (ICML 2025 · similarity 0.86)
- Anytime-valid, Bayes-assisted, Prediction-Powered Inference (NeurIPS 2025 · similarity 0.85)
- Regression for the Mean: Auto-Evaluation and Inference with Few Labels through Post-hoc Regression (ICML 2025 · similarity 0.85)
- Prediction-Powered Causal Inferences (NeurIPS 2025 · similarity 0.85)
- QuEst: Enhancing Estimates of Quantile-Based Distributional Measures Using Model Predictions (ICML 2025 · similarity 0.84)
- FAB-PPI: Frequentist, Assisted by Bayes, Prediction-Powered Inference (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.