arXiv subjects

Doudou Zhou

Publications and source records attributed to Doudou Zhou.

At least 19 recordsLinked to original sources

arXiv2026

Graph-Based Testing of Conditional Mean Independence with Applications to Large Language Model Evaluation

We study whether a predictor $X$ provides additional information for predicting a scalar response $Y$ after adjustment for other variables $Z$, formalized by testing $\mathsf{H}_0:\E(Y\mid X,Z)=\E(Y\mid Z)$. Existing methods often require estimating nuisance functions involving $X$, which can be difficult for high-dimensional or structured predictors. We propose a graph-based test that extends the edge count principle to covariate adjustment. It represents $X$ through pairwise dissimilarities, estimates only the baseline regression $\E(Y\mid Z)$, and aggregates products of held-out residuals across graph edges. We develop sample-splitting and two-fold cross-fitting versions. Under regularity conditions, both have standard normal null limits, yielding analytic $p$-values, and are consistent under graph-smooth and graph-rough alternatives. Simulations with vector, image, and network predictors show good size control and power advantages over competing procedures, with cross-fitting improving power across the main alternatives. Applications examine human quality ratings, mathematical error labels, and LLM response preferences. Our method detects associations with quality ratings and an LLM judge's preferences after adjustment for text length and other recorded characteristics.

stat.ME↗

arXiv2026

Learning Sequential Decisions from Multiple Sources via Group-Robust Markov Decision Processes

Learning a common policy from multiple source environments requires accounting for cross-source heterogeneity, and the offline setting introduces an additional challenge of limited data coverage. We formulate this problem as a distributionally robust Markov decision process under a shared linear MDP representation with source-specific transition and reward components. The uncertainty set is constructed through feature-wise mixtures of these components, yielding a $d$-rectangular structure that captures cross-source heterogeneity and admits a robust Bellman recursion. Building on this formulation, we develop Group-Robust Pessimistic Value Iteration (GR-PEVI), an offline policy-learning algorithm that combines feature-wise worst-case aggregation with a data-dependent pessimism penalty. Under a robust partial coverage condition, we establish finite-sample suboptimality guarantees for GR-PEVI and extend the analysis to clustered sources, characterizing the trade-off between the statistical gain from pooling similar sources and the approximation error induced by within-cluster heterogeneity. Simulation studies demonstrate the effectiveness of GR-PEVI, with empirical behavior consistent with the theoretical analysis. We apply GR-PEVI to a multi-source electronic health record study of multiple sclerosis treatment decisions to illustrate its use in a clinical setting.

stat.ME↗

arXiv2026

Simultaneous Inference for Centre-Outward Rankings of Random Objects

In many applications one object is chosen to represent a sample and the least typical objects are set aside, usually without any measure of uncertainty. We measure how central an object is by its reference depth, the probability that a random reference point is closer to it than to a random opponent. Using only pairwise dissimilarities, we construct simultaneous intervals for all ranks, a centre set that contains the most central observation under stated conditions, and outlier sets that contain only objects among the least central ones, each with asymptotic probability at least a prescribed level. These guarantees hold even though the ranked objects are themselves random. For symmetric distributions on the real line, the centre set has the smallest possible order of size up to a logarithmic factor, and in dimension three or more it reduces to the most central observation with probability tending to one. For typical meteorological years the centre set usually contains several candidate years, so the data do not determine the year on which building-energy simulations rest; for three collections of gene trees it contains one tree, whose features can be reported as those of the most central tree.

stat.ME↗

arXiv2026

AgentCue: Statistical Summaries for Guiding Decentralized Multi-Agent Self-Organization

Multi-agent systems (MAS) combine multiple agents to solve complex tasks, and recent MAS scale to more than a thousand agents. At this scale, decentralized MAS remove the central orchestrator that assigns sub-tasks, and the agents, which we call workers, self-organize through their own choices of sub-tasks and collaborators on a shared board. As workers keep adding participation and collaboration records, however, the information explodes: workers spend more and more effort on reading the records of other workers, and self-organization loses efficiency. In the statistical view, we represent the execution history of an MAS by two continually updated networks, a collaboration network among workers and a participation network between workers and task types. We then propose AgentCue, a method built on a new latent space model that places workers and task types in a shared latent space. Each time the networks are updated, the estimated parameters of the latent space model form a compact and interpretable statistical summary, and workers use the statistical summary to choose sub-tasks and collaborators without repeated searches of the shared board. We establish the identifiability of the latent space model and non-asymptotic error bounds for the estimates under network increments, and we show that a fixed number of gradient updates after each network update suffices for the same error rate. Experiments on LongBench v2 Multi-Doc QA show that AgentCue achieves competitive effectiveness, efficiency, and interpretability.

cs.MA↗

arXiv2026

Human-Anchored Inference for Ranking New Models with Large Language Model Judges

Human pairwise comparisons provide a reference for evaluating large language models (LLMs), but collecting sufficient judgments for each new release is costly and time-consuming. LLM judges offer a scalable alternative, although their comparisons may differ systematically from human preferences and across judges. We study the ranking of a new model that has received LLM-judge comparisons but no human comparisons. We propose ANCHOR (ANchored Comparisons for Human-reference inference with Orthogonal Riesz correction), which uses historical human and LLM comparisons to learn judge-specific sensitivities to human score differences and feature-dependent judge biases. These estimates are then used to infer the new model's human-reference score from its judge comparisons. The framework allows the feature distribution to change between historical and new-model comparisons. For inference, we construct a Neyman-orthogonal estimator through a joint Riesz correction that removes the first-order effects of estimating the historical human scores, judge sensitivities, and bias functions. We establish identification, convergence rates, and asymptotic normality with consistently estimable variance, and show that ANCHOR attains the semiparametric efficiency bound. Simulations demonstrate gains in score estimation and ranking accuracy. On Chatbot Arena, ANCHOR achieves the lowest score RMSE and insertion MAE among competing methods, with narrower score intervals on average.

math.ST↗

arXiv2026

Bad Genius: Counterfactual-Guided Harness Evolution Beyond Task-Specific Shortcuts

Reliable agent evaluation is complicated by automatic harness optimization, which repeatedly uses a released benchmark $B_{\mathrm{rel}}$ to guide a Proposer that edits prompts, memory, retrieval, tools, and control code around a fixed foundation model. Task holdout is commonly used to guard against harness overfitting. It varies semantic tasks but leaves the benchmark protocol fixed, so a bad genius Proposer can produce a cheating harness whose improvement over the initial harness on $B_{\mathrm{rel}}$ depends on a benchmark-wide shortcut. We introduce Counterfactual Harness Search and Evolution (CHASE), which casts harness evolution as constraint generation over valid counterfactual benchmarks. After each Proposer update, a Challenger searches for an executable protocol transformation with large gain destruction. A validity firewall checks that task semantics are preserved, while a held-out confirmation set determines whether the counterfactual enters a finite archive. We formalize an ideal shortcut-neutralized benchmark $B_0$ and establish theoretical guarantees linking finite counterfactual archives to $B_0$ and characterizing sequential Challenger search. We evaluate CHASE on Syn-Ledger and OfficeQA, where CHASE retains strong released-benchmark gains while substantially reducing gain destruction under valid protocol transformations.

cs.AI↗

arXiv2026

Shapley Value Estimation for Multi-Site Data with Blockwise-Missing Features

Shapley value (SV)-based methods are the prevailing framework for feature attribution in machine learning, yet existing population-level Shapley estimators generally assume that observations used to evaluate the coalitional game are fully observed under a common feature space. This assumption is routinely violated in multi-site studies across biomedicine, social science, and environmental monitoring, where institutions record different features under different protocols, producing systematic blockwise missingness across sources. We first show that the standard remedy of imputing missing features before computing Shapley values introduces systematic, coalition-dependent bias into the resulting attributions. We then propose \textbf{FUSHAP} (\textbf{Fu}sion \textbf{Sh}apley \textbf{A}ttribution from \textbf{P}artially-observed data), a method that leverages partially-observed auxiliary sites to reduce the variance of a preliminary single-site Shapley estimate without imputation. A permutation-based screening step detects and excludes sites whose data distributions are incompatible with the target population. In synthetic experiments, FUSHAP achieves $3$--$8\times$ lower MSE than the single-site estimator and $2$--$3\times$ lower MSE than imputation baselines without incurring imputation-induced bias, and the screening procedure identifies misaligned sites with $82\%$ power at moderate misalignment and $100\%$ for strong misalignment. On multi-site air quality and multi-center clinical data, FUSHAP reduces MSE by approximately $3$--$7\times$ relative to the single-site estimator; in the clinical application, standard imputation can increase MSE above the single-site baseline.

stat.ML↗

arXiv2026

Generalized Independence Test

Testing independence is a widely encountered problem in modern data analysis, especially in high-dimensional settings where complex dependency structures are common. Traditional methods often lack power in such scenarios. To address this, we propose a novel test statistic that leverages both similarity and dissimilarity information to capture intricate relationships within the data. The proposed test demonstrates strong power across a wide range of alternatives in high-dimensional settings, as shown through extensive simulation studies. Under mild conditions, we prove that the permutation null distribution of the statistic converges to the $χ^2_4$ distribution, enabling straightforward control of the type I error. Our theoretical analysis further advances the moment method to establish joint asymptotic normality for a class of double-indexed permutation statistics. Moreover, we prove the power consistency of the proposed test without imposing explicit restrictions on the data dimension, under mild conditions that allow for detecting complex dependencies. We illustrate the proposed test using gene-expression data from the Genotype-Tissue Expression project, where GIT rejects independence between the GTEx category "Breast - Mammary Tissue" and the cell-line category "Cells - EBV-transformed lymphocytes."

stat.ME↗

arXiv2026

Pattern-Calibrated Multimodal Prediction under Blockwise Missingness

Blockwise missingness in multimodal data is usually treated as an incomplete-input problem. We instead focus on prediction for a prespecified observed-modality pattern, where the observed modality set determines the information on which the prediction rule can condition. A procedure that imputes missing modalities, zero-fills unobserved modalities, or trains a single pooled predictor may borrow information across patterns, but it can also mix pattern-specific prediction rules. We propose Multimodal Overlap-aware Shared-specific Alignment and Inter-pattern Calibration (MOSAIC), a pattern-calibrated framework for borrowing across missingness patterns without collapsing their prediction rules. MOSAIC learns shared and modality-specific representations, uses the available representations that overlap with the target pattern to fit a first-stage predictor, and then estimates the calibration gap from target-pattern data. We establish non-asymptotic bounds that decompose the error into overlap effective sample size, calibration gap, and representation-learning error, clarifying when cross-pattern borrowing improves over local fitting and when the improvement is controlled by rule mismatch or representation-learning error. Simulations examine representation recovery and target-pattern correction, and applications to ICU mortality prediction, emotion recognition, and glaucoma classification show gains when target-pattern samples are limited or pattern-specific rules differ.

stat.ME↗

arXiv2026

Conditional Evaluation of Language Models with Cheap Auxiliary Signals

Aggregate accuracy hides where models succeed and fail. Estimating conditional performance profiles from gold labels alone is expensive, while cheap auxiliary signals such as LLM-judge scores, pairwise comparisons, confidence scores, and judge-disagreement features can be collected for every benchmark item but are often biased or miscalibrated. We propose LACE (Local Augmented Control-Variate Evaluation), a semi-supervised estimator for conditional LLM evaluation. The key step is local centering: after subtracting the conditional mean of a cheap signal within the target profile region, any linear augmentation has zero conditional mean and therefore cannot change the estimand. The augmentation coefficient is used only for efficiency, and a local ridge control variate combines a gold-label residual mean from the labeled subset with a cheap-signal mean from the full item pool. We prove calibration-free identification, unbiasedness for grouped profiles, local oracle optimality within centered linear augmentations, and first-order adaptivity to the estimated coefficient. The resulting gain formula is governed by a population local $R^2$, which characterizes how the efficiency attainable from the cheap signals varies across profile values. We also derive corresponding estimators for direct paired model gaps and deployment-weighted scores. We empirically evaluate the primary performance-profile estimator on MATH-500, ScienceQA, MMLU, WinoGrande, HellaSwag, TruthfulQA, GSM8K, and ARC.

cs.LG↗

arXiv2026

Inferential Evaluation of Surrogate-Derived Models under Covariate Shift

In transfer-learning settings, a model derived from abundant surrogate labels may be deployed in a target population where gold-standard outcomes are unobserved. Evaluating its target performance is essential for determining whether decisions based on the model remain reliable, yet it is difficult when gold labels are scarce, and covariate distributions differ across data sources. We study a three-sample setting with a small gold-labeled source, a larger surrogate-labeled source, and an unlabeled target. Under conditional transportability, we evaluate the surrogate-derived model against the latent gold-standard outcome in the target population. We propose cross-fitted estimators that transport information from the two labeled sources through source-specific density ratios. We also combine outcome-regression augmentation with a kernel correction for estimating the model near a threshold, accounting for uncertainty from all three samples. We establish asymptotically linear inference for TPR and FPR, consistency and pointwise inference for the ROC curve, and asymptotically normal inference for AUC. Simulations assess bias, coverage, and sensitivity to bandwidth and relative sample sizes. A retrospective temporal validation on Chatbot Arena and a semi-synthetic ACS-Income study provide validation in real-world AI applications.

stat.ML↗

arXiv2026

Generalized Linear Markov Decision Process

Offline reinforcement learning for longitudinal studies often faces two linked challenges: rewards may be binary or bounded, and reward observations may be available only for a subset of trajectories or time points even when the corresponding state-action-next-state histories are available. Linear Markov decision process methods are tractable because Bellman backups remain linear, but they require linear rewards and do not indicate how transition-only observations should be used. We introduce GRASP-MDP, Generalized Reward And Semi-supervised Pessimism for Markov Decision Processes, a reward-transition separated framework addressing both issues through one Bellman decomposition. It preserves linear transition dynamics while modeling reward means through generalized linear models. Although this breaks the usual linear Bellman form, the backup remains explicit as a nonlinear reward component plus a linear continuation component, yielding a nonlinear-plus-linear Bellman-complete class. In the resulting recursion, observed rewards estimate the generalized reward component, while all available transitions estimate the continuation component. The resulting pessimistic value iteration controls the two estimation errors separately without reward imputation. Finite-sample guarantees show that transition-only observations reduce transition-estimation error while reward uncertainty remains governed by observed rewards. Simulations and a multiple sclerosis electronic health record application illustrate the empirical benefit of retaining transition-only observations.

stat.ML↗

arXiv2026

Personalizing Large Language Model Agents with Small Policy Models

Large language model (LLM) agents can retrieve memory, call tools, ask clarifying questions, and vary response style, yet adapting these execution decisions to an individual user remains difficult. Fine-tuning a separate LLM is costly or impossible for proprietary systems, while prompts and memory primarily expose user information to the agent rather than adapt its execution decisions from feedback. We formulate personalization of a frozen agent as online learning of a per-user execution policy from scalar feedback observed only for the executed action. We propose FABLE (Factorized Adaptive Bandit Layer for Execution), a lightweight policy layer outside a potentially black-box host agent. FABLE factorizes memory, information-acquisition, and response decisions so feedback updates related choices; filters actions through an externally specified feasible set before exploration; and learns user-specific residual preferences relative to a fixed default-and-cost score via Bayesian contextual Thompson sampling. Under a linear residual-reward model, a calibrated variant inherits an expected-regret bound against the best feasible action. We also characterize preferences unidentifiable under persistent feasibility constraints and provide anytime-valid false-promotion control. Across personalized-reasoning, controlled-feedback, and executable tool-use evaluations, FABLE improves several preference-sensitive behaviors relative to rule-only control while remaining competitive on end-to-end task performance.

cs.AI↗

arXiv2026

Domain Adaptation Targeting Heterogeneous and Imbalanced Subgroups

Domain adaptation enables generalizable and efficient data-driven research. However, existing work has largely focused on domain adaptation for some intrinsically homogeneous target cohort, overlooking inherent heterogeneity within the target, which can exacerbate biases and unfairness in the presence of subgroups with imbalanced sample sizes. We develop a novel domain adaptation framework that addresses more complicated target data that consists of heterogeneous and data-sparse subgroups and lacks gold-standard label observations. Our method simultaneously handles high-dimensionality, covariate shift, and outcome model heterogeneity by combining a model-assisted debiasing step used for covariate shift correction with an adaptive knowledge-guided sparsification procedure used to mitigate the issue of sample disparity. We also introduce a new model selection strategy to avoid negative knowledge transfer in the absence of labels in the target data. Our method is theoretically justified for being robust to nuisance model misspecification and adaptive to heterogeneity between the subgroups. Numerical experiments and two real-world applications, including genetic risk modeling of type II diabetes and prediction of mutation-induced protein stability changes, demonstrate the practical advantages of our method.

stat.ME↗

arXiv2026

Contrastive Learning on Multimodal Analysis of Electronic Health Records

Electronic health record (EHR) systems capture a wealth of multimodal clinical data, encompassing both structured clinical codes and unstructured clinical notes. Yet, many EHR-focused studies have traditionally examined these modalities in isolation or combined them using simplistic methods, overlooking the intrinsic synergy between them. In reality, these modalities are deeply interconnected, each containing clinically relevant and complementary information that, when integrated effectively, can provide a more comprehensive understanding of patient health. Despite the success of multimodal contrastive learning in vision-language applications, its potential remains under-explored in multimodal EHR, particularly in terms of theoretical understanding. To support statistical analysis of multimodal EHR data, we propose a multimodal feature embedding generative model and design a multimodal contrastive loss to learn EHR feature representations. Our theoretical analysis demonstrates the effectiveness of multimodal learning over single-modality learning and connects the solution of the loss function to the singular value decomposition of a pointwise mutual information matrix. This connection leads to a privacy-preserving algorithm tailored for multimodal EHR representation learning. Simulation studies show that the proposed algorithm performs well under a variety of configurations. We further validate its clinical utility using real-world EHR data.

stat.ML↗

arXiv2026

Spherical Mixture Integration for Latent Embedding Alignment across Multi-Source Feature Spaces

Multi-institutional electronic health record (Multi-EHR) data have emerged as a powerful resource for developing predictive models to support clinical decisions and for generating reliable real-world evidence. By aggregating information from diverse patient populations and institutions, they enhance the robustness and generalizability of models and findings. However, analyzing multi-EHR remains challenging because disparate institutions rarely map all data elements to common ontologies, and raw EHR codes are often overly granular and institution-specific, fragmenting representations of the same clinical concept. Hence, integrative analysis must overcome two key hurdles: harmonizing codes with the same clinical meaning (synonymy), and aligning institutional feature spaces. To address these challenges, we propose SMILE, a Spherical Mixture Integration for Latent Embedding alignment across multi-source feature spaces, where embeddings from heterogeneous sources serve as privacy-preserving summaries of clinical concepts and sparse relational pairs provide weak supervision. Synonymy is modeled via a mixture of von Mises-Fisher distributions, yielding unified representations of semantically equivalent raw codes. We develop a composite quasi-likelihood estimator with non-asymptotic error bounds for the latent representations and mixture mean directions and consistent synonym-cluster recovery, quantifying the gains from integrating multiple sources and knowledge-graph information. Simulations and a multi-institutional EHR application demonstrate improved alignment and synonym clustering.

stat.ME↗

arXiv2026

A Judge-Aware Ranking Framework for Evaluating Large Language Models without Ground Truth

Evaluating large language models (LLMs) on open-ended tasks without ground-truth labels is increasingly done via the LLM-as-a-judge paradigm. A critical but under-modeled issue is that judge LLMs differ substantially in reliability; treating all judges equally can yield biased leaderboards and misleading uncertainty estimates. More data can make evaluation more confidently wrong under misspecified aggregation. We propose a judge-aware ranking framework that extends the Bradley-Terry-Luce model by introducing judge-specific discrimination parameters, jointly estimating latent model quality and judge reliability from pairwise comparisons without reference labels. We establish identifiability up to natural normalizations and prove consistency and asymptotic normality of the maximum likelihood estimator, enabling confidence intervals for score differences and rank comparisons. Across multiple public benchmarks and a newly collected dataset, our method improves agreement with human preferences, achieves higher data efficiency than unweighted baselines, and produces calibrated uncertainty quantification for LLM rankings.

stat.ML↗

arXiv2026

Structured Transfer Learning for Survival Risk Stratification in Data-Sparse Clinical Cohorts

Background: Survival prediction models are often less reliable in clinical groups with limited sample sizes or few outcome events. Target-only models may be unstable, whereas models from larger cohorts may transfer poorly when risk-factor effects differ across populations. We evaluated whether structured transfer learning can improve survival risk stratification in data-sparse cohorts while allowing cohort-specific adaptation. Methods: We developed the COhort-shared Rank-rEduced Cox model (CORE-Cox), a two-stage framework for multi-outcome survival prediction. CORE-Cox learns shared risk-factor patterns across related outcomes in a larger source cohort via a low-rank Cox coefficient structure, then adapts these patterns to a smaller target cohort through regularized residual correction. We evaluated CORE-Cox in UK Biobank (White source, n=150,093; Asian target, n=2,534) and MIMIC-IV (White ICU source, n=15,997; Asian ICU target, n=672), comparing against target-only Cox, penalized Cox, low-rank multi-task, naive pooling, direct transfer, and single-outcome residual transfer under repeated nested cross-validation. Results: CORE-Cox achieved best or near-best discrimination across most outcomes. Mean C-index improved from 0.733 to 0.766 in UK Biobank and from 0.628 to 0.658 in MIMIC-IV, with gains in eight of nine outcomes. CORE-Cox also improved top-15% risk enrichment, with hazard-ratio estimates typically intermediate between source-only and target-only models. Discussion: CORE-Cox offers an interpretable transfer-learning framework for survival risk stratification in data-sparse cohorts, combining shared cross-outcome structure with cohort-specific adaptation. Further validation is needed before use in calibrated absolute-risk prediction or clinical decision-making.

stat.ME↗