Xinrui Ruan | Searcharxiv

arXiv2026

Can Language Models Boost the Power of Randomized Experiments Without Statistical Bias?

Randomized controlled trials (RCTs) are widely adopted for causal inference, yet cost and sample-size constraints limit power. We introduce CALM (Causal Analysis leveraging Language Models), a statistical framework that integrates insights generated by large language models (LLMs) into the analysis of RCTs using established causal estimators to increase precision while preserving statistical validity. In particular, CALM treats LLM-generated outputs as auxiliary prognostic information and corrects their potential bias via a heterogeneous calibration step that residualizes and optimally reweights predictions. We prove that CALM remains consistent even when LLM predictions are biased and achieves efficiency gains over augmented inverse probability weighting estimators for various causal estimands. In particular, CALM develops a few-shot variant that aggregates predictions across randomly sampled demonstration sets. The resulting U-statistic-like predictor restores i.i.d. structure and also mitigates prompt-selection variability. Empirically, in simulations calibrated to a mobile-app depression RCT, CALM delivers lower variance relative to other benchmarking methods, is effective in zero- and few-shot settings, and remains stable across prompt designs. By principled use of LLMs to harness unstructured data and external knowledge learned during pretraining, CALM provides a practical path to more precise causal analyses.

stat.ME↗

arXiv2026

HERO: Improving the Reliability and Sensitivity of Generative Model Evaluation Using Historical Data

Reliable generative AI models critically rely on expert human annotations to evaluate output quality, yet these "gold" labels are expensive to collect and limited in quantity. Organizations thus often turn to collecting vast but noisy "silver" labels from crowdsourced workers or vendor annotators as proxies for gold labels. Because gold remains the evaluation target, naively aggregating noisy silver labels may introduce bias, and estimators built on sparsely observed gold labels may have high variance to resolve the model performance gaps that guide practical decisions. Model evaluation has become an ongoing operational practice rather than a one-time exercise, with evaluation rounds repeating across model versions, releases, and content domains. A natural question is whether the previous historical evaluation data can be used to improve each new round of evaluation. We introduce HERO (History Enhanced RObust model evaluation), a novel framework that uses historical data to suppress bias (improve reliability) and reduce variance (improve sensitivity) in model performance evaluation. HERO calibrates silver labelers' performance learned from historical gold annotations, and stabilizes the resulting estimator by anchoring it to covariate information measured with high precision in the historical data. HERO can be broadly applied across multiple common evaluation tasks, and remains valid when only a subset of historical labelers appears in the current round. We establish conditions under which the bias and variance reductions hold, showcase HERO's performance in simulation studies, and demonstrate its effectiveness on real-world model evaluation benchmarking datasets.

stat.ME↗