Extending Item Response Theory for Efficient and Meaningful Multilingual Evaluation

Multilingual benchmarks are central to evaluating large language models (LLMs) across languages, but they suffer from three issues: exhaustive evaluation scales linearly with the number of languages, automatic translation introduces errors that are easily missed at scale, and some items conflate general and culture-specific knowledge. We address all three with a unified statistical framework, Multilingual-IRT, which extends Item Response Theory with per-language difficulty deviations, split discriminability separating content from language effects, and per-language ability residuals. Fitting Multilingual-IRT on 25 LLMs across 29 languages of MMLU-Pro-X, we show that its fitted parameters support three practical applications: predicting unobserved (item, LLM, language) instances with 11-16% lower binary cross-entropy than the strongest accuracy-based baseline, surfacing candidate translation errors distributed across all 28 non-English languages, whereas accuracy-based baselines concentrate detections in a few languages, and recovering culture-specific items that accuracy-based baselines miss.

Robust CATE Estimation Using Novel Ensemble Methods

The estimation of Conditional Average Treatment Effects (CATE) is crucial for understanding the heterogeneity of treatment effects in clinical trials. We evaluate the performance of common methods, including causal forests and various meta-learners, across a diverse set of scenarios revealing that each of the methods fails in one or more of the tested scenarios. Given the inherent uncertainty of the data-generating process in real-life scenarios, the robustness of a CATE estimator to various scenarios is critical for its reliability. To address this limitation of existing methods, we propose two new ensemble methods that integrate multiple estimators to enhance prediction stability and performance - Stacked X-Learner which uses the X-Learner with model stacking for estimating the nuisance functions, and Consensus Based Averaging (CBA), which averages only the models with highest internal agreement. We show that these models achieve good performance across a wide range of scenarios varying in complexity, sample size and structure of the underlying-mechanism, including a biologically driven model for PD-L1 inhibition pathway for cancer treatment.

Causal Responder Detection

We introduce the causal responders detection (CARD), a novel method for responder analysis that identifies treated subjects who significantly respond to a treatment. Leveraging recent advances in conformal prediction, CARD employs machine learning techniques to accurately identify responders while controlling the false discovery rate in finite sample sizes. Additionally, we incorporate a propensity score adjustment to mitigate bias arising from non-random treatment allocation, enhancing the robustness of our method in observational settings. Simulation studies demonstrate that CARD effectively detects responders with high power in diverse scenarios.