Papers

Traditional LLM alignment methods are vulnerable to heterogeneity in human preferences. Fitting a naïve probabilistic model to pairwise comparison data (say over prompt-completion pairs) yields an inconsistent estimate of the population-average utility -a canonical measure of social welfare. We propose a new method, dubbed the sign estimator, that provides a simple, provably consistent, and efficient estimator by replacing cross-entropy with binary classification loss in the aggregation step. This simple modification recovers consistent ordinal alignment under mild assumptions and achieves the first polynomial finite-sample error bounds in this setting. In realistic simulations of LLM alignment using digital twins, the sign estimator substantially reduces preference distortion over a panel of simulated personas, cutting (angular) estimation error by nearly 35% and decreasing disagreement with true population preferences from 12% to 8% compared to standard RLHF. Our method also compares favorably to panel data heuristics that explicitly model user heterogeneity and require tracking individual-level preference data-all while maintaining the implementation simplicity of existing LLM alignment pipelines.

Designing frictional interfaces to exhibit prescribed macroscopic behavior is a challenging inverse problem, made difficult by the non-uniqueness of solutions and the computational cost of contact simulations. Traditional approaches rely on heuristic search over low-dimensional parameterizations, which limits their applicability to more complex or nonlinear friction laws. We introduce a generative modeling framework using Variational Autoencoders (VAEs) to infer surface topographies from target friction laws. Trained on a synthetic dataset composed of 200 million samples constructed from a parameterized contact mechanics model, the proposed method enables efficient, simulation-free generation of candidate topographies. We examine the potential and limitations of generative modeling for this inverse design task, focusing on balancing accuracy, throughput, and diversity in the generated solutions. Our results highlight trade-offs and outline practical considerations when balancing these objectives. This approach paves the way for near-real-time control of frictional behavior through tailored surface topographies.

Survivors of childhood cancer need lifelong monitoring for side effects from radiotherapy. However, longitudinal data from routine monitoring is often infrequently and irregularly sampled, and subject to inaccuracies. Due to this, measurements are often studied in isolation, or simple relationships (e.g., linear) are used to impute missing timepoints. In this study, we investigated the potential role of Gaussian Processes (GP) modelling to make population-based and individual predictions, using insulin-like growth factor 1 (IGF-1) measurements as a test case. With training data of 23 patients with a median (range) of 4 (1-16) timepoints we identified a trend within the range of literature reported values. In addition, with 8 test cases, individual predictions were made with an average root mean squared error of 31.9 (10.1 - 62.3) ng/ml and 27.4 (0.02 - 66.1) ng/ml for two approaches. GP modelling may overcome limitations of routine longitudinal data and facilitate analysis of late effects of radiotherapy.

Clinical decision support systems (CDSS) are used to improve clinical and service outcomes, yet evidence from low- and middle-income countries (LMICs) is dispersed. This protocol outlines methods to quantify the impact of CDSS on patient and healthcare delivery outcomes in LMICs. We will include comparative quantitative designs (randomized trials, controlled before-after, interrupted time series, comparative cohorts) evaluating CDSS in World Bank-defined LMICs. Standalone qualitative studies are excluded; mixed-methods studies are eligible only if they report comparative quantitative outcomes, for which we will extract the quantitative component. Searches (from inception to 30 September 2024) will cover MEDLINE, Embase, CINAHL, CENTRAL, Web of Science, Global Health, Scopus, IEEE Xplore, LILACS, African Index Medicus, and IndMED, plus grey sources. Screening and extraction will be performed in duplicate. Risk of bias will be assessed with RoB 2 (randomized trials) and ROBINS-I (non-randomized). Random-effects meta-analysis will be performed where outcomes are conceptually or statistically comparable; otherwise, a structured narrative synthesis will be presented. Heterogeneity will be explored using relative and absolute metrics and a priori subgroups or meta-regression (condition area, care level, CDSS type, readiness proxies, study design).

Machine learning models have achieved widespread success but often inherit and amplify historical biases, resulting in unfair outcomes. Traditional fairness methods typically impose constraints at the prediction level, without addressing underlying biases in data representations. In this work, we propose a principled framework that adjusts data representations to balance predictive utility and fairness. Using sufficient dimension reduction, we decompose the feature space into target-relevant, sensitive, and shared components, and control the fairness-utility trade-off by selectively removing sensitive information. We provide a theoretical analysis of how prediction error and fairness gaps evolve as shared subspaces are added, and employ influence functions to quantify their effects on the asymptotic behavior of parameter estimates. Experiments on both synthetic and real-world datasets validate our theoretical insights and show that the proposed method effectively improves fairness while preserving predictive performance.

Multi-category data arise in diverse fields including marketing, chemistry, public policy, genomics, political science, and ecology. We consider the problem of estimating ratios of category-specific means in a fully nonparametric setting, allowing for both observational units and categories to be preferentially sampled. We consider covariate-adjusted and unadjusted estimands that are non-parametrically defined and straightforward to interpret. While identifiability for related models has been established through parametric distributions or restrictions on the conditional mean (e.g., log-linearity), we show that identifiability can be obtained through an independence assumption or a category constraint, such as a reference category or a centering function. We develop an efficient, doubly-robust targeted minimum loss based estimator with excellent finite-sample performance, including in the setting of a large number of infrequently observed categories. We contrast the performance of our method with related approaches via simulation, and apply it to identify bacteria that are differentially abundant in diarrheal cases compared to controls. Our work provides a general framework for studying parameter identifiability in compositional data settings without requiring parametric assumptions on the data distribution.

Large language models are increasingly used to classify text at scale. Because their outputs are stochastic, it is common to query a model several times per item and aggregate the results, and latent class models offer a principled way to convert such repeated ratings into estimates of outcome prevalence and of the model's error rates. These models assume that repeated ratings of the same item are conditionally independent given the true label. We examine this assumption using 52,500 judgments from seven open-weight models ranging from 1B to 32B parameters, collected under three prompt conditions with ten samples per item on a public benchmark with human labels. On the balanced split of the benchmark, intraclass correlations lie between 0.50 and 0.93, so ten ratings carry the information of between 1.07 and 1.81 independent ratings. Under this correlation, a latent class model with a binomial likelihood is both miscalibrated and biased. Its nominal 95% intervals for the false acceptance rate contained the value computed from human labels in none of 42 settings, and it underestimated the total error rate in all 42. We propose a beta-binomial latent class model with a single correlation parameter, establish its identifiability and that of extensions to covariates, treatment effects and partially labeled data, and show that on the balanced split it raises the coverage of nominal 95% intervals for the prevalence and the false acceptance rate to 0.90. Recovery of the latent state is governed by the gap between the model's acceptance rates on truly positive and truly negative items, which equals Youden's J and predicts recovery almost monotonically (Spearman correlation 0.96). When this gap is small, a labeled subset improves estimates of prevalence but not the classification of individual items. Simulation studies designed around the identification conditions support these findings.

We propose a Bayesian variable selection method in the framework of modal regression for heavy-tailed responses. An efficient expectation-maximization algorithm is employed to expedite parameter estimation. A test statistic is constructed to exploit the shape of the model error distribution to effectively separate informative covariates from unimportant ones. Through simulations, we demonstrate and evaluate the efficacy of the proposed method in identifying important covariates in the presence of non-Gaussian model errors. Finally, we apply the proposed method to analyze two datasets arising in genetic and epigenetic studies.

This article studies statistical estimation of $\pi$ based on the fact that the ratio of the volumes of a $d$-dimensional hypersphere and a $d$-dimensional hypercube is a certain function of $\pi$, and the function depends on the dimension $d$. The estimation of $\pi$ is carried out for various choices of $d$ (strictly speaking, $d\in\{1, 2, \ldots, 20\}$) using the idea of Monte Carlo simulations. Various intriguing facts are observed, and the estimation of $\pi$ using infinite dimensional observations is outlined. Moreover, the R codes associated with relevant numerical studies are provided.

Calibration of mean estimates is a crucial property in many applications, particularly in actuarial decision-making. Tests of mean-calibration are often performed under the assumption that the estimated means correctly preserve the ordering of the corresponding population means. In this paper, we first consider the classical Likelihood Ratio Test (LRT) within the Exponential Dispersion Family (EDF), and we show that applying likelihood-based methods to test mean-calibration within the EDF requires a strong ordering assumption on the mean estimates. This assumption makes the null hypothesis of the mean-calibration test equivalent to the correct specification of the mean estimates. Next, we investigate the framework of universal inference for testing mean-calibration and mean-specification, and the relationship between the tests. We develop a sub-sampled split LRT within the EDF that provides finite-sample guarantees with universally valid critical values. We investigate type I error, power, and e-power of the sub-sampled split LRT and compare its performance with that of the classical LRT. We also propose a novel test statistic based on a sub-sampled split $L_q$-likelihood ratio test (L$q$RT) to enhance the performance of the statistical test. Numerical results demonstrate that universal inference based on the sub-sampled split LRT and sub-sampled split L$q$RT provides an attractive alternative to the classical LRT within the EDF, achieving high power for detecting miscalibration and misspecification in medium- and large-sized samples.

Industrial processes generate complex data that challenge fault detection systems, often yielding opaque or underwhelming results despite advanced machine learning techniques. This study tackles such difficulties using the Tennessee Eastman Process, a well-established benchmark known for its intricate dynamics, to develop an innovative fault detection framework. Initial attempts with standard models revealed limitations in both performance and interpretability, prompting a shift toward a more tractable approach. By employing SHAP (SHapley Additive exPlanations), we transform the problem into a more manageable and transparent form, pinpointing the most critical process features driving fault predictions. This reduction in complexity unlocks the ability to apply causal analysis through Directed Acyclic Graphs, generated by multiple algorithms, to uncover the underlying mechanisms of fault propagation. The resulting causal structures align strikingly with SHAP findings, consistently highlighting key process elements-like cooling and separation systems-as pivotal to fault development. Together, these methods not only enhance detection accuracy but also provide operators with clear, actionable insights into fault origins, a synergy that, to our knowledge, has not been previously explored in this context. This dual approach bridges predictive power with causal understanding, offering a robust tool for monitoring complex manufacturing environments and paving the way for smarter, more interpretable fault detection in industrial systems.

The predictive accuracy of operator learning frameworks depends on the quality and quantity of available training data (input-output function pairs), often requiring substantial amounts of high-fidelity data, which can be challenging to obtain in some real-world engineering applications. These datasets may be unevenly discretized from one realization to another, with the grid resolution varying across samples. In this study, we introduce a physics-informed operator learning approach by extending the Resolution Independent Neural Operator (RINO) framework to a fully data-free setup, addressing both challenges simultaneously. Here, the arbitrarily (but sufficiently finely) discretized input functions are projected onto a latent embedding space (i.e., a vector space of finite dimensions), using pre-trained basis functions. The operator associated with the underlying partial differential equations (PDEs) is then approximated by a simple multi-layer perceptron (MLP), which takes as input a latent code along with spatiotemporal coordinates to produce the solution in the physical space. The PDEs are enforced via a finite difference solver in the physical space. The validation and performance of the proposed method are benchmarked on several numerical examples with multi-resolution data, where input functions are sampled at varying resolutions, including both coarse and fine discretizations.

Purpose: Hereditary cancer risk is key to guiding screening and prevention strategies. Cancer risks can vary by individual due to the presence or absence of high- and moderate-risk pathogenic variants (PV) in cancer-associated genes, in addition to sex, age, and other risk factors. We previously developed Fam3PRO, a flexible multi-gene, multi-cancer Mendelian risk prediction model that estimates a patient's risk of carrying a PV in hereditary cancer genes and their future risk of developing several types of cancer. The Fam3PRO R package includes 22 genes with 18 associated cancers, allowing users to build customized sub-models from any gene-cancer set. However, the current R package lacks a user interface (UI), limiting its practical use in clinical settings. Therefore, we aim to develop a web-based UI for broader use of the Fam3PRO functionalities. Methods: The Fam3PRO UI (F3PI), built with R Shiny, collects and formats inputs including family health history, genetic test results, and other risk factors. Pedigree data are interactively visualized and modified via pedigreejs, while the backend Fam3PRO model takes all the inputs to generate carrier probabilities and future cancer risks, presented through an interactive UI. Results: F3PI streamlines the collection of patient and family history data, which is analyzed by the Fam3PRO models to provide personalized cancer risks for each proband across 18 cancers, as well as probabilities that a proband has a PV in up to 22 hereditary cancer genes. These results are returned to the user, within one minute on average and are available in both interactive and downloadable formats. Conclusion: We have developed F3PI, an easy-to-use, interactive web application that makes cancer and genetic risk information more accessible to providers and their patients.

Transitioning from Phase 2 to Phase 3 in drug development, at a rate of $\approx$40%, is the most stringent among phase transitions (Hay et al. (2014)). Yet, success rate at Phase 3 leading to approval is only $\approx$50% (Arrowsmith (2011b)). To improve Confirmability, we propose a methodological shift: replacing multiple hypothesis testing with inference based on confidence sets, and substituting conventional power and sample size calculations with a Confidently Bounded Quantile (CBQ) framework. Our confidence set inferences to answer the questions of whether to transition to a Confirmatory study as well as what to designate as the endpoint in that study. Construction of our directed confidence sets follows the Partitioning Principle, taking the best of each of Pivoting and Neyman Confidence Set Construction. Rooted in Tukey's Confidently Bounded Allowance (CBA) (Tukey (1994a)), our proposed CBQ makes the transitioning decision following the Correct and Useful Inference principle in Hsu (1996). CBQ removes from "power" the probability of rejecting for wrong reasons, eliminating the need for informal discounting in power calculation that has existed in the biopharmaceutical industry. ETZ, the modeling principle proposed in Wang et al. (2025), quantifies the impact of three variability components on confirmability. In repeated-measures RCTs, it separates within-subject and between-subject variability, further dividing the latter into baseline and trajectory components. This enables informed investment decisions for the sponsors on targeting variability reduction to improve confirmability. A Shiny-based Confirmability App supports all computations.

Alternating recurrent events, where subjects experience two potentially correlated event types over time, are common in healthcare, social, and behavioral studies. Often there is a primary event of interest that, when triggered, initiates a period of treatment and recovery measured via a secondary time-to-event. For example, cancer patients can experience repeated blood clotting emergencies that require hospitalization followed by discharge, people with alcohol use disorder can have periods of addiction and sobriety, or care partners can experience periods of depression and recovery. Potential censoring of the data requires special handling. Overlaying this are the missing at-risk periods for the primary event type when individuals have initiated the primary event but not reached the subsequent secondary event. In this paper, we develop a framework for regression analysis of censored alternating recurrent events that uses a random forest inverse probability weighting strategy to avoid bias in the analysis of the time to the primary event due to informative missingness from the alternate secondary state. The proposed regression model estimates $\tau$-restricted mean time to the primary event of interest while taking into account complexities of censored. Simulations show good performance of our method when the alternate times-to-event are either independent or correlated. We analyze a mobile health study data to evaluate the impact of self-care push notifications on the mental state of caregivers of traumatic brain injury patients.

Recent alignment techniques, such as reinforcement learning from human feedback, have been widely adopted to align large language models with human preferences by learning and leveraging reward models. In practice, these models often exploit spurious correlations, involving, e.g., response length, discrimination, sycophancy, and conceptual bias, which is a problem that has received increasing attention. In this work, we propose a principled framework that mitigates these biases in reward models while preserving the underlying factors that reflect intended preferences. We first provide a formulation of the data-generating process, assuming that the observed data (e.g., text) is generated from both spurious and non-spurious latent variables. We show that, interestingly, these non-spurious latent variables can be theoretically identified from data, regardless of whether a surrogate for the spurious latent variables is available. This further inspires a practical method that uses variational inference to recover these variables and leverages them to train reward models. Experiments on synthetic and real-world datasets demonstrate that our method effectively mitigates spurious correlation issues and yields more robust reward models.

Quantifying the uncertainty in the output of a neural network is essential for deployment in scientific or engineering applications where decisions must be made under limited or noisy data. Bayesian neural networks (BNNs) provide a framework for this purpose by constructing a Bayesian posterior distribution over the network parameters. However, the prior, which is of key importance in any Bayesian setting, is rarely meaningful for BNNs. This is because the complexity of the input-to-output map of a BNN makes it difficult to understand how certain distributions enforce any interpretable constraint on the output space of the network. Gaussian processes (GPs), on the other hand, are often preferred in uncertainty quantification tasks due to their interpretability. The drawback is that GPs are limited to small datasets without advanced techniques, which often rely on the covariance kernel having a specific structure. To address these challenges, we introduce a new class of priors for BNNs, called Mercer priors, such that the resulting BNN has samples which approximate that of a specified GP. The method works by defining a prior directly over the network parameters from the Mercer representation of the covariance kernel, and does not rely on the network having a specific structure. In doing so, we can exploit the scalability of BNNs in a meaningful Bayesian way.

Particle filters are a widely used Monte Carlo based data assimilation technique that estimates the probability distribution of a system's state conditioned on observations through a collection of weights and particles. A known problem for particle filters is weight collapse, or degeneracy, where a single weight attains a value of one while all others are close to zero, thereby collapsing the estimated distribution. We address this issue by introducing a novel modification to the particle filter that is simple to implement and inspired by energy-based diversity measures. Our approach adjusts particle weights to minimize a two-body energy potential, promoting balanced weight distributions and mitigating collapse. We demonstrate the performance of this modified particle filter in a series of numerical experiments with linear and nonlinear dynamical models, where we compare with the classical particle filter and ensemble Kalman filters in the nonlinear case. We find that our new approach improves weight distributions compared to the classical particle filter and thereby improve state estimates.

Deep neural networks (DNNs) have been used to model complex optimization problems in many applications, yet have difficulty guaranteeing solution optimality and feasibility, despite training on large datasets. Training a NN as a surrogate optimization solver amounts to estimating a global solution function that maps varying problem input parameters to the corresponding optimal solutions. Work in multiparametric programming (mp) has shown that solutions to quadratic programs (QP) are piece-wise linear functions of the parameters, and researchers have suggested leveraging this property to model mp-QP using NN with ReLU activation functions, which also exhibit piecewise linear behaviour. This paper proposes a NN modeling approach and learning algorithm that discovers the exact closed-form solution to QP with linear constraints, by analytically deriving NN model parameters directly from the problem coefficients without training. Whereas generic DNN cannot guarantee accuracy outside the training distribution, the closed-form NN model produces exact solutions for every discovered critical region of the solution function. To evaluate the closed-form NN model, it was applied to DC optimal power flow problems in electricity management. In terms of Karush-Kuhn-Tucker (KKT) optimality and feasibility of solutions, it outperformed a classically trained DNN and was competitive with, or outperformed, a commercial analytic solver (Gurobi) at far less computational cost. For a long-range energy planning problem, it was able to produce optimal and feasible solutions for millions of input parameters within seconds.

Variational mean field approximations tend to struggle with contemporary overparametrized deep neural networks. Where a Bayesian treatment is usually associated with high-quality predictions and uncertainties, the practical reality has been the opposite, with unstable training, poor predictive power, and subpar calibration. Building upon recent work on reparametrizations of neural networks, we propose a simple variational family that considers two independent linear subspaces of the parameter space. These represent functional changes inside and outside the support of training data. This allows us to build a fully-correlated approximate posterior reflecting the overparametrization that tunes easy-to-interpret hyperparameters. We develop scalable numerical routines that maximize the associated evidence lower bound (ELBO) and sample from the approximate posterior. Empirically, we observe state-of-the-art performance across tasks, models, and datasets compared to a wide array of baseline methods. Our results show that approximate Bayesian inference applied to deep neural networks is far from a lost cause when constructing inference mechanisms that reflect the geometry of reparametrizations.

In the current work, we provide theoretical results for testing (in)dependence between pairs of paths of most commonly studied non-stationary Gaussian processes - standard Brownian motion and fractional Brownian motion (fBm). Please see the PDF version of the paper for a full abstract.

Training or fine-tuning large language model (LLM)-based systems often requires costly human feedback, yet there is limited understanding of how to minimize such intervention while maintaining strong error guarantees. We study this problem for LLM-based classification systems in an active learning framework: an agent sequentially labels $d$-dimensional query embeddings drawn i.i.d. from an unknown distribution by either calling a costly expert or guessing with no feedback, with the goal of minimizing regret relative to an oracle with free expert access. When the horizon $T$ is at least exponential in the embedding dimension $d$, the geometry of the class regions can be learned. In this regime, we propose the Conservative Hull-based Classifier (CHC), which maintains convex hulls of expert-labeled queries and calls the expert when a query lands outside all known hulls. CHC attains $\mathcal{O}(\log^d T)$ regret in $T$ and is minimax optimal for $d=1$. Otherwise, the geometry cannot be reliably learned in general. We show that for queries drawn from a subgaussian mixture and $T \le e^d$, a Center-based Classifier (CC) achieves regret proportional to $N\log{N}$ where $N$ is the number of labels. To bridge these regimes, we introduce the Generalized Hull-based Classifier (GHC), a practical extension of CHC that enables more aggressive guessing via a tunable parameter. Our approach is validated on real-world question-answering datasets using state-of-the-art text embedding models.

The estimation of unknown parameters in nonlinear partial differential equations (PDEs) offers valuable insights across a wide range of scientific domains. In this work, we focus on estimating plant root parameters in the Richards equation, which is essential for understanding the soil-plant system in agricultural studies. Since conventional methods are computationally intensive and often yield unstable estimates, we develop a new Gaussian process collocation method for efficient Bayesian inference. Unlike existing Gaussian process-based approaches, our method constructs an approximate posterior distribution using samples drawn from a Gaussian process model fitted to the observed data, which does not require any structural assumption about the underlying PDE. Further, we propose to use an importance sampling procedure to correct for the discrepancy between the approximate and true posterior distributions. As an alternative, we also devise a prior-guided Bayesian optimization algorithm leveraging the approximate posterior. Simulation studies demonstrate that our method yields robust estimates under various settings. Finally, we apply our method on a real agricultural data set and estimate the plant root parameters with uncertainty quantification.

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.

Anonymizing microdata requires balancing the reduction of disclosure risk with the preservation of data utility. Traditional evaluations often rely on single measures or two-dimensional risk-utility (R-U) maps, but real-world assessments involve multiple, often correlated, indicators of both risk and utility. Pairwise comparisons of these measures can be inefficient and incomplete. We therefore systematically compare six visualization approaches for simultaneous evaluation of multiple risk and utility measures: heatmaps, dot plots, composite scatterplots, parallel coordinate plots, radial profile charts, and PCA-based biplots. We introduce blockwise PCA for composite scatterplots and joint PCA for biplots that simultaneously reveal method performance and measure interrelationships. Through systematic identification of Pareto-optimal methods in all approaches, we demonstrate how multivariate visualization supports a more informed selection of anonymization methods.

In this paper, we leverage stochastic projection and lossy compression to establish new conditional mutual information (CMI) bounds on the generalization error of statistical learning algorithms. It is shown that these bounds are generally tighter than the existing ones. In particular, we prove that for certain problem instances for which existing MI and CMI bounds were recently shown in Attias et al. [2024] and Livni [2023] to become vacuous or fail to describe the right generalization behavior, our bounds yield suitable generalization guarantees of the order of $\mathcal{O}(1/\sqrt{n})$, where $n$ is the size of the training dataset. Furthermore, we use our bounds to investigate the problem of data "memorization" raised in those works, and which asserts that there are learning problem instances for which any learning algorithm that has good prediction there exist distributions under which the algorithm must "memorize" a big fraction of the training dataset. We show that for every learning algorithm, there exists an auxiliary algorithm that does not memorize and which yields comparable generalization error for any data distribution. In part, this shows that memorization is not necessary for good generalization.

AI predictive systems are increasingly embedded in decision making pipelines, shaping high stakes choices once made solely by humans. Yet robust decisions under uncertainty still rely on capabilities that current AI lacks: domain knowledge not captured by data, long horizon context, and reasoning grounded in the physical world. This gap has motivated growing efforts to design collaborative frameworks that combine the complementary strengths of humans and AI. This work advances this vision by identifying the fundamental principles of Human AI collaboration within uncertainty quantification, a key component of reliable decision making. We introduce Human AI Collaborative Uncertainty Quantification, a framework that formalizes how an AI model can refine a human expert's proposed prediction set with two goals: avoiding counterfactual harm, ensuring the AI does not degrade correct human judgments, and complementarity, enabling recovery of correct outcomes the human missed. At the population level, we show that the optimal collaborative prediction set follows an intuitive two threshold structure over a single score function, extending a classical result in conformal prediction. Building on this insight, we develop practical offline and online calibration algorithms with provable distribution free finite sample guarantees. The online method adapts to distribution shifts, including human behavior evolving through interaction with AI, a phenomenon we call Human to AI Adaptation. Experiments across image classification, regression, and text based medical decision making show that collaborative prediction sets consistently outperform either agent alone, achieving higher coverage and smaller set sizes across various conditions.

Calibration has emerged as a foundational goal in “trustworthy machine learning”, in part because of its strong decision theoretic semantics. Independent of the underlying distribution, and independent of the decision maker's utility function, calibration promises that amongst all policies mapping predictions to actions, the uniformly best policy is the one that “trusts the predictions” and acts as if they were correct. But this is true only of fully calibrated forecasts, which are tractable to guarantee only for very low dimensional prediction problems. For higher dimensional prediction problems (e.g. when outcomes are multiclass), weaker forms of calibration have been studied that lack these decision theoretic properties. In this paper we study how a conservative decision maker should map predictions endowed with these weaker (“partial”) calibration guarantees to actions, in a way that is robust in a minimax sense: i.e. to maximize their expected utility in the worst case over distributions consistent with the calibration guarantees. We characterize their minimax optimal decision rule via a duality argument, and show that surprisingly, “trusting the predictions and acting accordingly” is recovered in this minimax sense by decision calibration (and any strictly stronger notion of calibration), a substantially weaker and more tractable condition than full calibration. For calibration guarantees that fall short of decision calibration, the minimax optimal decision rule is still efficiently computable, and we provide an empirical evaluation of a natural one that applies to any regression model solved to optimize squared error.

Federated Learning (FL) is a distributed learning paradigm that preserves privacy by eliminating the need to exchange raw data during training. In its prototypical edge instantiation with underlying wireless transmissions enabled by analog over-the-air computing (AirComp), referred to as over-the-air FL (AirFL), the inherent channel noise plays a unique role of frenemy in the sense that it degrades training due to noisy global aggregation while providing a natural source of randomness for privacy-preserving mechanisms, formally quantified by differential privacy (DP). It remains, nevertheless, challenging to effectively harness such channel impairments, as prior arts, under assumptions of either simple channel models or restricted types of loss functions, mostly considering (local) DP enhancement with a single-round or non-convergent bound on privacy loss. In this paper, we study AirFL over multiple-access fading channels with a multi-antenna base station (BS) subject to user-level DP requirements. Despite a recent study, which claimed in similar settings that artificial noise (AN) must be injected to ensure DP in general, we demonstrate, on the contrary, that DP can be gained as a perk even without employing any AN. Specifically, we derive a novel bound on DP that converges under general bounded-domain assumptions on model parameters, along with a convergence bound with general smooth and non-convex loss functions. Next, we optimize over receive beamforming and power allocations to characterize the optimal convergence-privacy trade-offs, which also reveal explicit conditions in which DP is achievable without compromising training. Finally, our theoretical findings are validated by extensive numerical results.

In this work, we study out-of-distribution (OOD) generalization in meta-reinforcement learning from an information-theoretic perspective. We begin by establishing OOD generalization bounds for meta-supervised learning under two distinct distribution shift scenarios: standard distribution mismatch and a broad-to-narrow training setting. Building on this foundation, we formalize the generalization problem in meta-reinforcement learning and establish fine-grained generalization bounds that exploit the structure of Markov Decision Processes. Lastly, we analyze the generalization performance of a gradient-based meta-reinforcement learning algorithm.