Randomized experiments are the preferred approach for evaluating the effects of interventions, but they are costly and often yield estimates with substantial uncertainty. On the other hand, in silico experiments leveraging foundation models offer a cost-effective alternative that can potentially attain higher statistical precision. However, the benefits of in silico experiments come with a significant risk: statistical inferences are not valid if the models fail to accurately predict experimental responses to interventions. In this paper, we propose a novel approach that integrates the predictions from multiple foundation models with experimental data while preserving valid statistical inference. Our estimator is consistent and asymptotically normal, with asymptotic variance no larger than the standard estimator based on experimental data alone. Importantly, these statistical properties hold even when model predictions are arbitrarily biased. Empirical results across several randomized experiments show that our estimator offers substantial precision gains, equivalent to a reduction of up to 20% in the sample size needed to match the same precision as the standard estimator based on experimental data alone.

Synthetic tabular data are often evaluated by distributional similarity, privacy distance, or train-on-synthetic-test-on-real predictive performance, but these criteria do not ensure validity for causal inference. We show that fully generative tabular synthesizers, including GAN- and LLM-based models, can preserve predictive utility while distorting average treatment effect (ATE) estimates. The failure is structural: ATE preservation requires both a realistic covariate law and an accurate treatment-effect contrast, whereas prediction loss penalizes treatment-effect error only through an overlap-weighted term. We formalize this mismatch through sensitivity and loss-decomposition results, and identify an analogous decomposition in block-level next-token prediction under log loss. Motivated by the tabular causal analysis, we propose a hybrid synthetic-data framework that generates covariates while modeling treatment and outcome mechanisms separately, allowing causal-purpose treatment assignment such as randomized synthetic assignment. We evaluate this framework in three settings: ATE preservation under fully generative versus hybrid synthesis, targeted augmentation for practical positivity problems, and synthetic simulation engines for comparing OR, IPW, AIPW, and TMLE before real-data analysis. Across synthetic and ACTG experiments, hybrid synthesis improves causal fidelity relative to fully generative baselines; LLM-based hybrid synthesis is often more faithful than CTGAN for ATE preservation and finite-sample estimator benchmarking.

Estimation of heterogeneous long-term treatment effects (HLTEs) is widely used for personalized decision-making in marketing, economics, and medicine, where short-term randomized experiments are often combined with long-term observational data. However, HLTE estimation is challenging due to limited overlap in treatment or in observing long-term outcomes for certain subpopulations, which can lead to unstable HLTE estimates with large finite-sample variance. To address this challenge, we introduce the LT-O-learners (Long-Term Orthogonal Learners), a set of novel orthogonal learners for HLTE estimation. The learners are designed for the canonical HLTE setting that combines a short-term randomized dataset $\mathcal{D}_1$ with a long-term historical dataset $\mathcal{D}_2$. The key idea of our LT-O-Learners is to retarget the learning objective by introducing custom overlap weights that downweight samples with low overlap in treatment or in long-term observation. We show that the retargeted loss is equivalent to the weighted oracle loss and satisfies Neyman-orthogonality, which means our learners are robust to errors in the nuisance estimation. We further provide a general error bound for the LT-O-Learners and give the conditions under which quasi-oracle rate can be achieved. Finally, our LT-O-learners are model-agnostic and can thus be instantiated with arbitrary machine learning models. We conduct empirical evaluations on synthetic and semi-synthetic benchmarks to confirm the theoretical properties of our LT-O-Learners, especially the robustness in low-overlap settings. To the best of our knowledge, ours are the first orthogonal learners for HLTE estimation that are robust to low overlap that is common in long-term outcomes.

We develop a direct debiased machine learning framework comprising Neyman targeted estimation and generalized Riesz regression. Our framework unifies Riesz regression for automatic debiased machine learning, covariate balancing, targeted maximum likelihood estimation (TMLE), and density-ratio estimation. In many problems involving causal effects or structural models, the parameters of interest depend on regression functions. Plugging regression functions estimated by machine learning methods into the identifying equations can yield poor performance because of first-stage bias. To reduce such bias, debiased machine learning employs Neyman orthogonal estimating equations. Debiased machine learning typically requires estimation of the Riesz representer and the regression function. For this problem, we develop a direct debiased machine learning framework with an end-to-end algorithm. We formulate estimation of the nuisance parameters, the regression function and the Riesz representer, as minimizing the discrepancy between Neyman orthogonal scores computed with known and unknown nuisance parameters, which we refer to as Neyman targeted estimation. Neyman targeted estimation includes Riesz representer estimation, and we measure discrepancies using the Bregman divergence. The Bregman divergence encompasses various loss functions as special cases, where the squared loss yields Riesz regression and the Kullback-Leibler divergence yields entropy balancing. We refer to this Riesz representer estimation as generalized Riesz regression. Neyman targeted estimation also yields TMLE as a special case for regression function estimation. Furthermore, for specific pairs of models and Riesz representer estimation methods, we can automatically obtain the covariate balancing property without explicitly solving the covariate balancing objective.

We study a web-deployed, tool-augmented LLM health coach with real users. In a pilot with seven users (280 rated turns), offline policy evaluation (OPE) over factorized decision heads (Tool/Style) shows that a uniform heavy-tool policy raises average value on logs but harms specific subgroups, most notably low-health-literacy/high-self-efficacy users. A lightweight simulator with hidden archetypes further shows that adding a small early information-gain bonus reliably shortens trait identification and improves goal success and pass@3. Together, these early findings indicate an evaluation-first path to personalization: freeze the generator, learn subgroup-aware decision heads on typed rewards (objective tool outcomes and satisfaction), and always report per-archetype metrics to surface subgroup harms that averages obscure.

Causal inference typically assumes centralized access to individual-level data. Yet, in practice, data are often decentralized across multiple sites, making centralization infeasible due to privacy, logistical, or legal constraints. We address this problem by estimating the Average Treatment Effect (ATE) from decentralized observational data via a Federated Learning (FL) approach, allowing inference through the exchange of aggregate statistics rather than individual-level data. We propose a novel method to estimate propensity scores by computing a federated weighted average of local scores with Membership Weights (MW)--probabilities of site membership conditional on covariates--which can be flexibly estimated using parametric or non-parametric classification models. Unlike density ratio weights (DW) from the transportability and generalization literature, which either rely on strong modeling assumptions or cannot be implemented in FL, MW can be estimated using standard FL algorithms and are more robust, as they support flexible, non-parametric models--making them the preferred choice in multi-site settings with strict data-sharing constraints. The resulting propensity scores are used to construct Federated Inverse Propensity Weighting (Fed-IPW) and Augmented IPW (Fed-AIPW) estimators. Unlike meta-analysis methods, which fail when any site violates positivity, our approach leverages heterogeneity in treatment assignment across sites to improve overlap. We show that Fed-IPW and Fed-AIPW perform well under site-level heterogeneity in sample sizes, treatment mechanisms, and covariate distributions. Both theoretical analysis and experiments on simulated and real-world data highlight their advantages over meta-analysis and related methods.

Large language model (LLM) evaluation is increasingly costly, prompting interest in methods that speed up evaluation by shrinking benchmark datasets. Benchmark prediction (also called efficient LLM evaluation) aims to select a small subset of evaluation points and predict overall benchmark performance from that subset. In this paper, we systematically assess the strengths and limitations of 11 benchmark prediction methods across 19 diverse benchmarks. First, we identify a highly competitive baseline: Take a random sample and fit a regression model on the sample to predict missing entries. Outperforming most existing methods, this baseline challenges the assumption that careful subset selection is necessary for benchmark prediction. Second, we discover that all existing methods crucially depend on model similarity. They work best when interpolating scores among similar models. The effectiveness of benchmark prediction sharply declines when new models have higher accuracy than previously seen models. In this setting of extrapolation, none of the previous methods consistently beat a simple average over random samples. To improve over the sample average, we introduce a new method inspired by augmented inverse propensity weighting. This method consistently outperforms the random sample average even for extrapolation. However, its performance still relies on model similarity and the gains are modest in general. This shows that benchmark prediction fails just when it is most needed: at the evaluation frontier, where the goal is to evaluate new models of unknown capabilities.

Estimating causal effects from observational data is essential in fields such as medicine, economics and social sciences, where privacy concerns are paramount. We propose a general, model-agnostic framework for differentially private estimation of average treatment effects (ATE) that avoids strong structural assumptions on the data-generating process or the models used to estimate propensity scores and conditional outcomes. In contrast to prior work, which enforces differential privacy by directly privatizing these nuisance components and results in a privacy cost that scales with model complexity, our approach decouples nuisance estimation from privacy protection. This separation allows the use of flexible, state-of-the-art black-box models, while differential privacy is achieved by perturbing only predictions and aggregation steps within a fold-splitting scheme with ensemble techniques. We instantiate the framework for three classical estimators -- the G-formula, inverse propensity weighting (IPW), and augmented IPW (AIPW) -- and provide formal utility and privacy guarantees. Empirical results show that our methods maintain competitive performance under realistic privacy budgets. We further extend our framework to support meta-analysis of multiple private ATE estimates. Our results bridge a critical gap between causal inference and privacy-preserving data analysis.

Constructing confidence intervals (CIs) for the average treatment effect (ATE) from patient records is crucial to assess the effectiveness and safety of drugs. However, patient records typically come from different hospitals, thus raising the question of how multiple observational datasets can be effectively combined for this purpose. In our paper, we propose a new method that estimates the ATE from multiple observational datasets and provides valid CIs. Our method makes little assumptions about the observational datasets and is thus widely applicable in medical practice. The key idea of our method is that we leverage predictionpowered inferences and thereby essentially'shrink' the CIs so that we offer more precise uncertainty quantification as compared to naïve approaches. We further prove the unbiasedness of our method and the validity of our CIs. We confirm our theoretical results through various numerical experiments. Finally, we provide an extension of our method for constructing CIs from combinations of experimental and observational datasets. Estimating the average treatment effect (ATE) together with confidence intervals (CIs) is relevant in many fields, such as medicine, where the ATE is used to assess the effectiveness and safety of drugs (Glass et al., 2013; Feuerriegel et al., 2024). Nowadays, there is a growing interest in using observational datasets for this purpose, for example, electronic health records (EHRs) and clinical registries (Johnson et al., 2016; Corrigan-Curay et al., 2018; Hong, 2021). Importantly, such observational datasets typically originate from different hospitals, different health providers, or even different countries (Colnet et al., 2024), thus raising the question of how to construct CIs for ATE estimation from multiple observational datasets. Motivating example: During the COVID-19 pandemic, the effectiveness and safety of potential drugs and vaccines were often assessed from electronic health records that originated from different hospitals to rapidly generate new evidence with treatment guidelines (Tacconelli et al., 2022). For example, one study (Wong et al., 2024) estimated the effect of nirmatrelvir/ritonavir (also known under the commercial name "paxlovid") in patients with COVID-19 diagnosis on 28-day all-cause hospitalizations from data obtained through a retrospective, multi-center study.