NLP models often rely on human-labeled data for training and evaluation. Many approaches crowdsource this data from a large number of annotators with varying skills, backgrounds, and motivations, resulting in conflicting annotations. These conflicts have traditionally been resolved by aggregation methods that assume disagreements are errors. Recent work has argued that for many tasks annotators may have genuine disagreements and that variation should be treated as signal rather than noise. However, few models separate signal and noise in annotator disagreement. In this work, we introduce NUTMEG, a new Bayesian model that incorporates information about annotator backgrounds to remove noisy annotations from human-labeled training data while preserving systematic disagreements. Using synthetic data, we show that NUTMEG is more effective at recovering ground-truth from annotations with systematic disagreement than traditional aggregation methods. We provide further analysis characterizing how differences in subpopulation sizes, rates of disagreement, and rates of spam affect the performance of our model. Finally, we demonstrate that downstream models trained on NUTMEG-aggregated data significantly outperform models trained on data from traditionally aggregation methods. Our results highlight the importance of accounting for both annotator competence and systematic disagreements when training on human-labeled data.

Background: Hypertensive kidney disease (HKD) patients in intensive care units (ICUs) face high short-term mortality, but tailored risk prediction tools are lacking. Early identification of high-risk individuals is crucial for clinical decision-making. Methods: We developed a machine learning framework to predict 30-day in-hospital mortality among ICU patients with HKD using early clinical data from the MIMIC-IV v2.2 database. A cohort of 1,366 adults was curated with strict criteria, excluding malignancy cases. Eighteen clinical features-including vital signs, labs, comorbidities, and therapies-were selected via random forest importance and mutual information filtering. Several models were trained and compared with stratified five-fold cross-validation; CatBoost demonstrated the best performance. Results: CatBoost achieved an AUROC of 0.88 on the independent test set, with sensitivity of 0.811 and specificity of 0.798. SHAP values and Accumulated Local Effects (ALE) plots showed the model relied on meaningful predictors such as altered consciousness, vasopressor use, and coagulation status. Additionally, the DREAM algorithm was integrated to estimate patient-specific posterior risk distributions, allowing clinicians to assess both predicted mortality and its uncertainty. Conclusions: We present an interpretable machine learning pipeline for early, real-time risk assessment in ICU patients with HKD. By combining high predictive performance with uncertainty quantification, our model supports individualized triage and transparent clinical decisions. This approach shows promise for clinical deployment and merits external validation in broader critical care populations.

Learning representations unaffected by superficial characteristics is important to ensure that shifts in these characteristics at test time do not compromise downstream prediction performance. For instance, in healthcare applications, we might like to learn features that contain information about pathology yet are unaffected by race, sex, and other sources of physiologic variability, thereby ensuring predictions are equitable and generalizable across all demographics. Here we propose Contrastive LEarning with Anti-contrastive Regularization (CLEAR), an intuitive and easy-to-implement framework that effectively separates essential (i.e., task-relevant) characteristics from superficial (i.e., task-irrelevant) characteristics during training, leading to better performance when superficial characteristics shift at test time. We begin by supposing that data representations can be semantically separated into task-relevant content features, which contain information relevant to downstream tasks, and task-irrelevant style features, which encompass superficial attributes that are irrelevant to these tasks, yet may degrade performance due to associations with content present in training data that do not generalize. We then prove that our anti-contrastive penalty, which we call Pair-Switching (PS), minimizes the Mutual Information between the style attributes and content labels. Finally, we instantiate CLEAR in the latent space of a Variational Auto-Encoder (VAE), then perform experiments to quantitatively and qualitatively evaluate the resulting CLEAR-VAE over several image datasets. Our results show that CLEAR-VAE allows us to: (a) swap and interpolate content and style between any pair of samples, and (b) improve downstream classification performance in the presence of previously unseen combinations of content and style. Our code will be made publicly available.

While diffusion models excel at image generation, their growing adoption raises critical concerns around copyright issues and model transparency. Existing attribution methods identify training examples influencing an entire image, but fall short in isolating contributions to specific elements, such as styles or objects, that matter most to stakeholders. To bridge this gap, we introduce \emph{concept-level attribution} via a novel method called \emph{Concept-TRAK}. Concept-TRAK extends influence functions with two key innovations: (1) a reformulated diffusion training loss based on diffusion posterior sampling, enabling robust, sample-specific attribution; and (2) a concept-aware reward function that emphasizes semantic relevance. We evaluate Concept-TRAK on the AbC benchmark, showing substantial improvements over prior methods. Through diverse case studies--ranging from identifying IP-protected and unsafe content to analyzing prompt engineering and compositional learning--we demonstrate how concept-level attribution yields actionable insights for responsible generative AI development and governance.

Publicly releasing the specification of a model with its trained parameters means an adversary can attempt to reconstruct information about the training data via training data reconstruction attacks, a major vulnerability of modern machine learning methods. This paper makes three primary contributions: establishing a mathematical framework to express the problem, characterising the features of the training data that are vulnerable via a maximum mean discrepancy equivalance and outlining a score matching framework for reconstructing data in both Bayesian and non-Bayesian models, the former is a first in the literature.

Accurate uncertainty quantification remains a key challenge for standard LLMs, prompting the adoption of Bayesian and ensemble-based methods. However, such methods typically necessitate computationally expensive sampling, involving multiple forward passes to effectively estimate predictive uncertainty. In this paper, we introduce a novel approach enabling efficient and effective uncertainty estimation in LLMs without sacrificing performance. Specifically, we distill uncertainty-aware teacher models - originally requiring multiple forward passes - into compact student models sharing the same architecture but fine-tuned using Low-Rank Adaptation (LoRA). We compare two distinct distillation strategies: one in which the student employs traditional softmax-based outputs, and another in which the student leverages Dirichlet-distributed outputs to explicitly model epistemic uncertainty via evidential learning. Empirical evaluations on classification datasets demonstrate that such students can achieve comparable or superior predictive and uncertainty quantification performance relative to their teacher models, while critically requiring only a single forward pass. To our knowledge, this is the first demonstration that immediate and robust uncertainty quantification can be achieved in LLMs through evidential distillation.

Given a multivariate function taking deterministic and unc ertain inputs, we consider the problem of estimating a quantile set: a set of deterministic inputs f or which the probability that the output belongs to a specific region remains below a given threshold. To solve this problem in the context of expensive-to-evaluate black-box functions, we propose a Bayesian active learning strategy based on Gaussian process modeling. The strategy is driven by a nov el sampling criterion, which belongs to a broader principle that we refer to as Expected Estimator Modification (EEM). More specifically, the strategy relies on a novel sampling criterion combined w ith a sequential Monte Carlo framework that enables the construction of batch-sequential designs for the efficient estimation of small quantile sets. The performance of the strategy is illustrated on seve ral synthetic examples and an industrial application case involving the ROTOR37 compressor model.

We introduce the microclustering Ewens--Pitman model for random partitions, obtained by scaling the strength parameter of the Ewens--Pitman model linearly with the sample size. The resulting random partition is shown to have the microclustering property, namely: the size of the largest cluster grows sub-linearly with the sample size, while the number of clusters grows linearly. By leveraging the interplay between the Ewens--Pitman random partition with the Pitman--Yor process, we develop efficient variational inference schemes for posterior computation in entity resolution. Our approach achieves a speed-up of three orders of magnitude over existing Bayesian methods for entity resolution, while maintaining competitive empirical performance.

Do larger and more capable language models learn to update their "beliefs" about propositions more consistently with Bayes' theorem when presented with evidence in-context? To test this, we formulate a Bayesian Coherence Coefficient (BCC) metric and generate a dataset with which to measure the BCC. We measure BCC for multiple pre-trained-only language models across five model families, comparing against the number of model parameters, the amount of training data, and model scores on common benchmarks. Our results provide evidence for our hypothesis that larger and more capable pre-trained language models assign credences that are more coherent with Bayes' theorem. These results have important implications for our understanding and governance of LLMs.

The most effective differentially private machine learning algorithms in practice rely on an additional source of purportedly public data. This paradigm is most interesting when the two sources combine to be more than the sum of their parts. However, there are settings such as mean estimation where we have strong lower bounds, showing that when the two data sources have the same distribution, there is no complementary value to combining the two data sources. In this work we extend the known lower bounds for public-private learning to setting where the two data sources exhibit significant distribution shift. Our results apply to both Gaussian mean estimation where the two distributions have different means, and to Gaussian linear regression where the two distributions exhibit parameter shift. We find that when the shift is small (relative to the desired accuracy), either public or private data must be sufficiently abundant to estimate the private parameter. Conversely, when the shift is large, public data provides no benefit.