UAI 2025 Papers

To effectively study complex causal systems, it is often useful to construct abstractions of parts of the system by discarding irrelevant details while preserving key features. The Information Bottleneck (IB) method is a widely used approach to construct variable abstractions by compressing random variables while retaining predictive power over a target variable. Traditional methods like IB are purely statistical and ignore underlying causal structures, making them ill-suited for causal tasks. We propose the Causal Information Bottleneck (CIB), a causal extension of the IB, which compresses a set of chosen variables while maintaining causal control over a target variable. This method produces abstractions of (sets of) variables which are causally interpretable, give us insight about the interactions between the abstracted variables and the target variable, and can be used when reasoning about interventions. We present experimental results demonstrating that the learned abstractions accurately capture causal relations as intended.

Training overparameterized neural networks often yields solutions with varying generalization capabilities, even when achieving similar training losses. Recent evidence indicates a strong correlation between the sharpness of a minimum and its generalization error, leading to increased interest in optimization methods that explicitly seek flatter minima for improved generalization. Despite its contemporary relevance to overparameterization, however, this sharpness-aware minimization (SAM) strategy has not been studied much yet as to exactly how it is affected by overparameterization. In this work, we analyze SAM under varying degrees of overparameterization, presenting both empirical and theoretical findings that reveal its critical influence on SAM’s effectiveness. First, we conduct extensive numerical experiments across diverse domains and show that SAM consistently improves with overparameterization. Next, we attribute this phenomenon to the interplay between the enlarged solution space and increased implicit bias resulting from overparameterization. Furthermore, we show that this effect is particularly pronounced in practical settings involving label noise and sparsity, and yet, sufficient regularization is necessary. Last but not least, we provide other theoretical insights into how overparameterization helps SAM achieve minima with more uniform Hessian moments compared to SGD, and much faster convergence at a linear rate.

We develop minimax optimal estimators for nonparametric regression methods when the true regression function lies in an L_2-fractional Sobolev space with order s∈ (0,1). This function class is a Hilbert space lying between the space of square-integrable functions and the first-order Sobolev space consisting of differentiable functions. It contains fractional power functions, piecewise constant or piecewise polynomial functions and bump function as canonical examples. We construct an estimator based on performing Principal Component Regression using Fractional Laplacian Eigenmaps and show that the in-sample mean-squared estimation error of this estimator is of order n^-2s/2s+d, where d is the dimension, s is the order parameter and n is the number of observations. We next prove a minimax lower bound of the same order, thereby establishing that no other estimator can improve upon the proposed estimator, up to context factors. We also provide preliminary empirical results validating the practical performance of the developed estimators.

Methods to improve the adversarial robustness of tree-based ensemble models for classification tasks have received significant attention in recent years. In this work, we propose a novel method for training robust tree-based boosted ensembles applicable to any task that employs a differentiable loss function, leveraging the XGBoost framework. Our work introduces an analytical solution to the upper-bound of the robust loss function, that can be computed in constant time, enabling the construction of robust splits without sacrificing computational efficiency. Although our method is general, we focus its application on regression tasks, extending conventional regression metrics to better quantify model robustness. An extensive evaluation on 19 regression datasets from a widely-used tabular data benchmark demonstrates that in the face of adversarial perturbations in the input space, our proposed method results in ensembles that are up to 44

Randomized Controlled Trials (RCTs) are a fundamental aspect of data-driven decision-making. RCTs often assume that the units are not influenced by each other. Traditional approaches addressing such effects assume a fixed network structure between the interfering units. However, real-world networks are rarely static, and treatment assignments can actively reshape the interference structure itself, as seen in financial access interventions that alter informal lending networks or healthcare programs that modify peer influence dynamics. This creates a novel and unexplored problem: estimating treatment effects when the interference network is determined by treatment allocation. In this work, we address this gap by proposing two single-experiment estimators for scenarios where network edges depend on nodal treatments constructed from instrumental variables derived from neighbourhood treatments. We prove their unbiasedness and experimentally validate the proposed estimators both on synthetic and real data.

Continual learning (CL) has gained increasing interest in recent years due to the need for models that can continuously learn new tasks while retaining knowledge from previous ones. However, existing CL methods often require either computationally expensive layer-wise gradient projections or large-scale storage of past task data, making them impractical for resource-constrained scenarios. To address these challenges, we propose a local model space projection (LMSP)-based continual learning framework that significantly reduces computational complexity from 𝒪(n^3) to 𝒪(n^2) while preserving both forward and backward knowledge transfer with minimal performance trade-offs. We establish a theoretical analysis of the error and convergence properties of LMSP compared to conventional global approaches. Extensive experiments on multiple public datasets demonstrate that our method achieves competitive performance while offering substantial efficiency gains, making it a promising solution for scalable continual learning.

We utilize Hybrid Markov Logic Networks (HMLNs) to combine embeddings learned from a Deep Neural Network (DNN) with symbolic relational knowledge. Since a DNN may not always learn optimal embeddings, we develop a mixture model to reduce variance in the HMLN parameterization. Further, we perform inference in our model that is robust to covariate shifts that may occur in the DNN embeddings by reparameterizing the HMLN. We evaluate our approach on Graph Neural Networks and show that our approach outperforms state-of-the-art methods that combine relational knowledge with DNN embeddings when we introduce covariate shifts in the embeddings. Further, we demonstrate the utility of our approach in inferring latent student knowledge in a cognitive model called Deep Knowledge Tracing.

We consider the problem of generating valid and small prediction sets by sampling outputs (e.g., software code and natural language text) from a black-box deep generative model for a given input (e.g., textual prompt). The validity of a prediction set is determined by a user-defined binary admissibility function depending on the target application. For example, requiring at least one program in the set to pass all test cases in code generation application. To address this problem, we develop a simple and effective conformal inference algorithm referred to as Generative Prediction Sets (GPS). Given a set of calibration examples and black-box access to a deep generative model, GPS can generate prediction sets with provable guarantees. The key insight behind GPS is to exploit the inherent structure within the distribution over the minimum number of samples needed to obtain an admissible output to develop a simple conformal regression approach over the minimum number of samples. Unlike prior work , the sets generated by GPS do not require iterative sampling at test time, while maintaining strict marginal coverage guarantees. Experiments on multiple datasets for code and math word problems using different large language models demonstrate the efficacy of GPS over state-of-the-art methods.

Probabilistic circuits (PCs) enable us to learn joint distributions over a set of random variables and to perform various probabilistic queries in a tractable fashion. Though the tractability property allows PCs to scale beyond non-tractable models such as Bayesian Networks, scaling training and inference of PCs to larger, real-world datasets remains challenging. To remedy the situation, we show how PCs can be learned across multiple machines by recursively partitioning a distributed dataset, thereby unveiling a deep connection between PCs and federated learning (FL). This leads to federated circuits (FCs)—a novel and flexible federated learning (FL) framework that (1) allows one to scale PCs on distributed learning environments (2) train PCs faster and (3) unifies for the first time horizontal, vertical, and hybrid FL in one framework by re-framing FL as a density estimation problem over distributed datasets. We demonstrate FC’s capability to scale PCs on various large-scale datasets. Also, we show FC’s versatility in handling horizontal, vertical, and hybrid FL within a unified framework on multiple classification tasks.

Gaussian Processes (GPs) provide a convenient framework for specifying function-space priors, making them a natural choice for modeling uncertainty. In contrast, Bayesian Neural Networks (BNNs) offer greater scalability and extendability but lack the advantageous properties of GPs. This motivates the development of BNNs capable of replicating GP-like behavior. However, existing solutions are either limited to specific GP kernels or rely on heuristics. We demonstrate that trainable activations are crucial for effective mapping of GP priors to wide BNNs. Specifically, we leverage the closed-form 2-Wasserstein distance for efficient gradient-based optimization of reparameterized priors and activations. Beyond learned activations, we also introduce trainable periodic activations that ensure global stationarity by design, and functional priors conditioned on GP hyperparameters to allow efficient model selection. Empirically, our method consistently outperforms existing approaches or matches performance of the heuristic methods, while offering stronger theoretical foundations.

Adversarially robust optimization (ARO) has emerged as the *de facto* standard for training models that hedge against adversarial attacks in the test stage. While these models are robust against adversarial attacks, they tend to suffer severely from overfitting. To address this issue, some successful methods replace the empirical distribution in the training stage with alternatives including *(i)* a worst-case distribution residing in an ambiguity set, resulting in a distributionally robust (DR) counterpart of ARO; *(ii)* a mixture of the empirical distribution with a distribution induced by an auxiliary (*e.g.*, synthetic, external, out-of-domain) dataset. Inspired by the former, we study the Wasserstein DR counterpart of ARO for logistic regression and show it admits a tractable convex optimization reformulation. Adopting the latter setting, we revise the DR approach by intersecting its ambiguity set with another ambiguity set built using the auxiliary dataset, which offers a significant improvement whenever the Wasserstein distance between the data generating and auxiliary distributions can be estimated. We study the underlying optimization problem, develop efficient solution algorithms, and demonstrate that the proposed method outperforms benchmark approaches on standard datasets.

Reliable estimation of predictive uncertainty is crucial for machine learning applications, particularly in high-stakes scenarios where hedging against risks is essential. Despite its significance, there is no universal agreement on how to best quantify predictive uncertainty. In this work, we revisit core concepts to propose a framework for information-theoretic measures of predictive uncertainty. Our proposed framework categorizes predictive uncertainty measures according to two factors: (I) The predicting model (II) The approximation of the true predictive distribution. Examining all possible combinations of these two factors, we derive a set of predictive uncertainty measures that includes both known and newly introduced ones. We extensively evaluate these measures across a broad set of tasks, identifying conditions under which certain measures excel. Our findings show the importance of aligning the choice of uncertainty measure with the predicting model on in-distribution (ID) data, the limitations of epistemic uncertainty measures for out-of-distribution (OOD) data, and that the disentanglement between measures varies substantially between ID and OOD data. Together, these insights provide a more comprehensive understanding of predictive uncertainty measures, revealing their implicit assumptions and relationships.

Despite their widespread use, large language models (LLMs) are known to hallucinate incorrect information and be poorly calibrated. This makes the uncertainty quantification of these models of critical importance, especially in high-stakes domains, such as autonomy and healthcare. Prior work has made Bayesian deep learning-based approaches to this problem more tractable by performing inference over the low-rank adaptation (LoRA) parameters of a fine-tuned model. While effective, these approaches struggle to scale to larger LLMs due to requiring further additional parameters compared to LoRA. In this work we present Scalable Bayesian Low-Rank Adaptation via Stochastic Variational Subspace Inference (ScalaBL). We perform Bayesian inference in an r-dimensional subspace, for LoRA rank r. By repurposing the LoRA parameters as projection matrices, we are able to map samples from this subspace into the full weight space of the LLM. This allows us to learn all the parameters of our approach using stochastic variational inference. Despite the low dimensionality of our subspace, we are able to achieve competitive performance with state-of-the-art approaches while only requiring ∼1000 additional parameters. Furthermore, it allows us to scale up to the largest Bayesian LLM to date, with four times as a many base parameters as prior work.

Comparing datasets is a fundamental task in machine learning, essential for various learning paradigms-from evaluating train and test datasets for model generalization to using dataset similarity for detecting data drift. While traditional notions of dataset distances offer principled measures of similarity, their utility has largely been assessed through prediction error minimization. However, in Predict-then-Optimize (PtO) frameworks, where predictions serve as inputs for downstream optimization tasks, model performance is measured through decision regret rather than prediction error. In this work, we propose OTD^3 (Optimal Transport Decision-aware Dataset Distance), a novel dataset distance that incorporates downstream decisions in addition to features and labels. We show that traditional feature-label distances lack informativeness in PtO settings, while OTD^3 more effectively captures adaptation success. We also derive a PtO-specific adaptation bound based on this distance. Empirically, we show that our proposed distance accurately predicts model transferability across three different PtO tasks from the literature. Code is available at https://github.com/paularodr/OTD3

We characterize a notion of confidence that arises in learning or updating beliefs: the amount of trust one has in incoming information and its impact on the belief state. This *learner’s confidence* can be used alongside (and is easily mistaken for) probability or likelihood, but it is fundamentally a different concept—one that captures many familiar concepts in the literature, including learning rates and number of training epochs, Shafer’s weight of evidence, and Kalman gain. We formally axiomatize what it means to learn with confidence, give two canonical ways of measuring confidence on a continuum, and prove that confidence can always be represented in this way. Under additional assumptions, we derive more compact representations of confidence-based learning in terms of vector fields and loss functions. These representations induce an extended language of compound "parallel" observations. We characterize *Bayesian* learning as the special case of an optimizing learner whose loss representation is a linear expectation.

In LLM alignment and many other ML applications, one often faces the *Multi-Objective Fine-Tuning* (MOFT) problem, *i.e.*, fine-tuning an existing model with datasets labeled w.r.t. different objectives simultaneously. To address the challenge, we propose a *Conditioned One-Shot* fine-tuning framework (COS-DPO) that extends the Direct Preference Optimization technique, originally developed for efficient LLM alignment with preference data, to accommodate the MOFT settings. By direct conditioning on the weight across auxiliary objectives, our Weight-COS-DPO method enjoys an efficient one-shot training process for profiling the Pareto front and is capable of achieving comprehensive trade-off solutions even in the post-training stage. Based on our theoretical findings on the linear transformation properties of the loss function, we further propose the Temperature-COS-DPO method that augments the temperature parameter to the model input, enhancing the flexibility of post-training control over the trade-offs between the main and auxiliary objectives. We demonstrate the effectiveness and efficiency of the COS-DPO framework through its applications to various tasks, including the Learning-to-Rank (LTR) and LLM alignment tasks, highlighting its viability for large-scale ML deployments.

In many real-world settings, such as medicine and finance causal effect is a valuable metric for decision making. For many predictive tasks, causal mechanisms provide robust estimators while existing ML-driven predictors might be vulnerable to spurious correlations. In such settings, when data is decentralized and privacy must be preserved, federated learning plays an important role. However, causal inference in a federated learning setup is a largely unexplored research area. In this paper, we learn a proxy of the underlying structural causal model (SCM) with deep generative models from decentralized observational data sources possibly containing high-dimensional variables. Based on client preference or high dimensionality of variables, we modularize the SCM mechanisms and find the minimal subset appropriate for federated learning while having rest of the mechanisms trained on individual client’s local data. When all connected together, the proxy SCM, named as the federated deep causal generative model (FeDCM ), offers estimation of any identifiable causal effect. We perform extensive experiments to illustrate the utility and performance of our approach.

In ex ante coordinated adversarial team games (ATGs), a team competes against an adversary, and team members can only coordinate their strategies before the game starts. The team-maxmin equilibrium with correlation (TMECor) is a suitable solution concept for extensive-form sequential ATGs. One class of TMECor-solving methods transforms the problem into solving NE in two-player zero-sum games, leveraging well-established tools for the latter. However, existing methods are fundamentally action-based, resulting in poor generalizability and low solving efficiency due to the exponential growth in the size of the transformed game. To address the above issues, we propose an efficient game transformation method based on private information, where all team members are represented by a single coordinator. We designed a structure called private information pre-branch, which makes decisions considering all possible private information from teammates. We prove that the size of the game transformed by our method is exponentially reduced compared to the current state-of-the-art. Moreover, we demonstrate equilibria equivalence. Experimentally, our method achieves a significant speedup of 182.89× to 694.44× in scenarios where the current state-of-the-art method can work, such as small-scale Kuhn poker and Leduc poker. Furthermore, our method is applicable to larger games and those with dynamically changing private information, such as Goofspiel.

The primary challenge of multi-label active learning, differing it from multi-class active learning, lies in assessing the informativeness of an indefinite number of labels while also accounting for the inherited label correlation. Existing studies either require substantial computational resources to leverage correlations or fail to fully explore label dependencies. Additionally, real-world scenarios often require addressing intrinsic biases stemming from imbalanced data distributions. In this paper, we propose a new multi-label active learning strategy to address both challenges. Our method incorporates progressively updated positive and negative correlation matrices to capture co-occurrence and disjoint relationships within the label space of annotated samples, enabling a holistic assessment of uncertainty rather than treating labels as isolated elements. Furthermore, alongside diversity, our model employs ensemble pseudo labeling and beta scoring rules to address data imbalances. Extensive experiments on four realistic datasets demonstrate that our strategy consistently achieves more reliable and superior performance, compared to several established methods.