UAI 2026 Papers

Dependent censoring occurs when the event time and censoring time are not conditionally independent given the observed covariates. This complicates survival model evaluation because widely used metrics, such as the Brier score, typically handle right-censoring using inverse probability of censoring weighting (IPCW). Unfortunately, IPCW is valid only when the estimated censoring distribution is independent of the event time. We propose a dependent Brier score based on an Archimedean copula and the Copula-Graphic estimator, and establish consistency and asymptotic normality of its margin-time estimator. To evaluate the metric, we introduce a semi-synthetic framework that creates realistic dependent censoring while preserving the original covariate structure and known event times. Across 12 datasets, the proposed metric reduces estimation error by 12-16

*Grokking*—delayed generalization that emerges well after a model has fit the training data—has been linked to robustness and representation quality. We ask whether this training regime also helps with *machine unlearning*, i.e., removing the influence of specified data without full retraining. We compare applying standard unlearning methods *before* versus *after* the grokking transition across vision (CNNs/ResNets on CIFAR, SVHN and ImageNet) and language (a transformer on a TOFU-style setup). Starting from grokked checkpoints consistently yields (i) more **efficient forgetting** (fewer updates to reach a target forget level), (ii) **less collateral damage** (smaller drops on retained and test performance), and (iii) **more stable updates** across seeds, relative to early-stopped counterparts under identical unlearning algorithms. Analyses of features and curvature further suggest that post-grokking models learn *more modular representations* with reduced gradient alignment between forget and retain subsets, which facilitates selective forgetting. Our results highlight **when** a model is trained (pre- vs. post-grokking) as an orthogonal lever to **how** unlearning is performed, providing a practical recipe to improve existing unlearning methods without altering their algorithms.

Cooperative multi-agent reinforcement learning (MARL) faces significant coordination challenges due to partial observability. Although communication can mitigate these issues, traditional methods often overlook the fact that agents can simultaneously belong to multiple collaborative groups, playing diverse roles. This limitation hinders the effective integration of tactical and strategic information. To address this, we propose Hierarchical Hypergraph Communication (HHC). In HHC, agents are modeled via a tactical-level hypergraph that supports overlapping multi-group memberships, facilitating the extraction of fine-grained tactical features. Simultaneously, these tactical hyperedges serve as virtual nodes to integrate and communicate strategic intents within a strategic-level hypergraph. This dual-layered architecture empowers agents to align individual actions with global strategies through a hierarchical communication process, thereby enhancing decision-making capabilities. Extensive experiments demonstrate the effectiveness of HHC in complex coordination tasks.

The problem of probabilistic verification of a neural network investigates the probability of satisfying the safe constraints in the output space when the input is given by a probability distribution. It is significant to answer this problem when the input is affected by disturbances which are often modeled by probabilistic variables. In the paper, we propose a novel neural network probabilistic verification framework which computes a guaranteed range for the safe probability by efficiently finding safe and unsafe probabilistic hulls. Our approach consists of three main innovations: (1) a state space subdivision strategy using regression trees to produce probabilistic hulls, (2) a boundary-aware sampling method which identifies the safety boundary in the input space using samples that are later used for building regression trees, and (3) iterative refinement with probabilistic prioritization for computing a guaranteed range for the safe probability. The accuracy and efficiency of our approach are evaluated on various benchmarks including ACAS Xu and a rocket lander controller. The result shows a clear improvement over the state of the art.

The contextual bandit models sequential decision-making problems in which rewards depend on both the chosen action and observed context. In many domains such as medicine, business, and engineering, prior knowledge provides structural information about the reward function that can be exploited to improve optimization efficiency. This paper studies settings where, for each observable context, the conditional mean reward is known to be concave with respect to the action variable. To leverage this structure, we develop a contextual bandit algorithm that conditions a Bayesian Gaussian Process posterior on concavity constraints. We propose a novel reward model that combines a concavity-preserving regression spline basis with a constrained GP posterior, yielding a tractable shape-constrained estimator. Building on this model, we construct a UCB algorithm and establish new posterior concentration inequalities for the constrained posterior, leading to regret guarantees that are never worse than applying standard GP-UCB without concavity information and can be strictly tighter when concavity is informative. Experiments on benchmark problems and a Warfarin dosing test application demonstrate substantial reductions in cumulative regret relative to state-of-the-art baselines.

Coarsening-based training for graph neural networks (GNNs), i.e. training on coarsened graphs rather than the original large ones, has become a promising direction for scaling GNNs to massive graphs. However, prior work has been evaluated almost exclusively on *homophilic* graphs, leaving the more challenging *heterophilic* settings underexplored. We show, both empirically and theoretically, that existing coarsening-based training methods suffer significant performance degradation on heterophilic graphs due to inevitable loss of graph information during coarsening. To address this, we propose **A**daptive **C**omplementary **E**nhancement, a plug-and-play, model-agnostic strategy that reintegrates the information discarded in coarsening: ACE learns a projector for re-constructing original node features and applies *anisotropic structural regularization* to embed local heterophily. We further adopt *homoscedastic uncertainty weighting* to adaptively balance the combined training objective of primary coarsened-graph training loss and full-graph auxiliary loss with augmented node features re-constructed by the heterophily-aware projector. Extensive experiments show that ACE drives consistent gains on heterophilic benchmarks while preserving competitive results on homophilic graphs with minimal computational overhead. Code is available at the GitHub repository: < https://github.com/vasile-paskardlgm/ACE >.

Currently, the way deep learning models recognize uncertainty remains inconsistent with human perception. In multi-label classification, quantifying uncertainty at the label level presents challenges, as each label may exhibit distinct model confidence levels. Understanding and decomposing label-specific uncertainty is essential for interpreting model behavior and ensuring reliable predictions. We build a hierarchical Bayesian methodology for multi-label classification that leverages a Type II likelihood and Empirical Bayes. Then we estimate and decompose label-wise uncertainties by the bias-variance decomposition. Our approaches offer four main contributions: (1) Type II likelihood maximization is data likelihood centric; (2) it can decompose label-wise uncertainty into the model variance, the model bias and data noise; (3) our uncertainty represented by model bias is intuitively interpretable when combined with observational data; and (4) when applied to out-of-distribution (OOD) detection task, it achieves a 6.88

In multi-agent reinforcement learning, communication is essential for effective cooperation. Most existing methods assume instantaneous message delivery, and the few delay-aware studies consider only external fixed or stochastic delays, neglecting queueing delays caused by competition for a shared channel. We formulate the shared channel as a queueing system with state-dependent service rates, and propose Congestion-Aware Intent Communication (CAIC). To ensure message validity under delay, CAIC employs a temporal masked autoencoder to predict each agent’s future trajectory and encode it into a delay-robust intent message, with attention-based fusion integrating asynchronous messages at the receiver. To maintain timeliness, CAIC dynamically adjusts communication frequency based on intent message changes and delay estimates, proactively mitigating congestion. Experiments on Hallway, MPE, and SMAC demonstrate that CAIC outperforms existing baselines on most evaluated scenarios in both no-delay and queueing-delay settings, and ablation studies reveal that stale messages under delay can be even more harmful than no communication.

Accurate prediction of drug-target affinity (DTA) is crucial for accelerating drug discovery. Existing methods struggle to model global long-range dependencies of sequences at linear complexity, and multi-layer perceptrons relying on fixed activation functions in graph neural networks face limitations in fitting complex non-linear topologies. We propose SGFF-DTA, a sequence-graph fusion framework based on bidirectional Mamba (BiMamba) and Fourier-Kolmogorov-Arnold Networks (Fourier-KAN). The framework employs BiMamba to capture global contextual semantics of sequences at linear complexity through bidirectional selective state space scanning. It integrates Fourier-KAN into graph neural networks to model high-order topological interactions using learnable non-linear transformations in the spectral space. To achieve cross-modal semantic alignment, we introduce pre-trained feature transfer and a cross-gated fusion module. On three benchmark datasets, SGFF-DTA significantly outperforms state-of-the-art methods in terms of mean squared error, demonstrating robust generalization capabilities under cold-start settings. Visual analysis further confirms model interpretability in accurately locating key binding sites and pharmacophores.

We study the problem of allocating indivisible items to a group of agents as these items arrive over time. Our focus is on scenarios where the utility of the items is weighted based on the waiting time since the last item was received, which is a demand-side time-dependent utility model. To ensure fairness, we concentrate on two criteria: temporal envy-freeness (TEF) and temporal proportionality (TProp). These criteria require that the allocation remains fair after each item is allocated. When there are two agents, we demonstrate the sufficient and necessary conditions of the weight function that ensure the existence of a TEF1 or TProp1 allocation. Additionally, we provide two polynomial-time algorithms that output a TEF2 or TProp2 allocation, with some restrictions on the weight function. When there are more than two agents, we show the hardness of determining the existence of TEF1 allocations. We also provide polynomial algorithms returning TEF1/TProp1 allocations with some specific valuation profiles. When there are multiple items per round, we also provide a polynomial algorithm returning TEF1/TProp1 allocations.

Causal discovery seeks to uncover causal relations from data, typically represented as causal graphs, and is essential for predicting the effects of interventions. While expert knowledge is required to construct principled causal graphs, many statistical methods have been proposed to leverage observational data with varying formal guarantees. Causal Assumption-based Argumentation (ABA) is a framework that uses symbolic reasoning to ensure correspondence between input constraints and output graphs, while offering a principled way to combine data and expertise. We explore the use of large language models (LLMs) as imperfect experts, eliciting semantic structural constraints from variable names and descriptions and integrating them with statistical evidence through Causal ABA. We propose ABAPC-LLM as a principled and conservative approach to hybrid data- and LLM-driven causal discovery. Additionally, we introduce an evaluation protocol to mitigate memorisation bias when assessing LLMs for causal discovery and show competitive performance on novel semantically grounded random benchmarks, as well as on standard small- and medium-sized benchmarks.

Matched-pair experimental designs aim to detect treatment effects by pairing participants and comparing within-pair outcome differences. In many situations, the overall effect size across the entire population is small. Then, the focus naturally shifts to identifying and targeting high treatment-effect regions where the intervention is most effective. This paper proposes a matched-pair experimental design that sequentially and actively enrolls patients in high treatment-effect regions. Importantly, we frame the identification of the target region as a classification problem and propose an active learning framework tailored to matched-pair designs. Our design not only reduces the experimental cost of detecting treatment efficacy, but also ensures that the identified regions enclose the entire high-treatment-effect regions. Our theoretical analysis of the framework’s label complexity and experiments in practical scenarios demonstrate the efficiency and advantages of the approach.

The Shapley value is a widely used concept in attribution problems, as it uniquely satisfies the axioms of linearity, consistency, equal treatment, and efficiency. Often, the inclusion AUC metric is used to evaluate the quality of player rankings, in order to identify positively participating players. However, it can be established that the Shapley value is not always reliable for this purpose. The core issue lies in its linearity: the Shapley value acts as a linear operator with an excessively large null space, which is likely to contain non-negligible perturbations that remain indistinguishable to the operator. To address this limitation, we explore the design of nonlinear axiomatic attribution methods. Inspired by the least core, which is a popular nonlinear substitute for the Shapley value, we introduce a class of nonlinear attribution methods that retain the remaining necessary axioms. Each method yields a contribution vector that is the unique optimal solution to a minimization problem, which aims to approximate utility functions as faithfully as possible. In terms of the inclusion AUC metric, our experiments demonstrate the potential effectiveness of these methods compared to Shapley value variants that relax only the efficiency axiom.

In this paper, we investigate tradeoffs among differential privacy (DP) and several representative axioms for approval-based committee voting, including justified representation, proportional justified representation, extended justified representation, Pareto efficiency, and Condorcet criterion. Without surprise, we demonstrate that all of these axioms are incompatible with DP, and thus establish both upper and lower bounds for their two-way tradeoffs with DP. Furthermore, we provide upper and lower bounds for three-way tradeoffs among DP and every pairwise combination of such axioms, revealing that although these axioms are compatible without DP, their optimal levels under DP cannot be simultaneously achieved. Our results quantify the effect of DP on the satisfaction and compatibility of the axioms in approval-based committee voting, which can provide insights for designing voting rules that possess both privacy and axiomatic properties.

Simple Temporal Networks with Uncertainty (STNUs) are a well-established formalism for reasoning about temporal plans involving uncontrollable durations, known as contingent links. This model enables checking the controllability of plans under different assumptions about when uncertainties are revealed (Weak, Dynamic, and Strong Controllability). Recent work has also introduced methods to repair uncontrollable STNUs by adjusting contingent bounds, which is relevant e.g., in multi-agent systems, or resource-aware settings where some external flexibility can be negotiated. Nevertheless, there remains a lack of formal understanding regarding the computational complexity of such repair problems. This paper fills that gap by formally defining the problems and assessing the complexity of a number of them. In particular, we show that for a number of settings, repair is no harder than controllability.

In this paper, we establish Berry-Esseen-type bounds for federated linear stochastic approximation (LSA). Our results provide the first federated Gaussian approximations for LSA that explicitly capture communication-computation trade-offs and heterogeneity-aware error terms, quantifying the effects of local step size, number of local updates, and heterogeneity on convergence rates. We present results for both (i) constant step size regime and (ii) decreasing step size with an increasing number of local iterations, recovering the recent rates of Bonnerjee et al. [2026] as a special case. As a primary application of our results, we develop an online multiplier bootstrap procedure for inference on the last iterate, which avoids explicit estimation of the asymptotic covariance matrix, and obtain non-asymptotic validity guarantees for this procedure.

We present a new data structure, *block cover*, for approximate weighted counting and sampling of subsets of a given query set. Such queries are the main computational bottleneck, e.g., in advanced Markov chain Monte Carlo algorithms for Bayesian learning of Bayesian networks. Given a collection of weighted subsets of a ground set N, our key idea is to select a few moderate-size subsets B of N, called blocks, so as to cover all or most of the input collection by the power sets 2^B. An approximate sum over the subsets of a query set Q is obtained by adding up the contributions within each block, which contibutions we precompute. We also give a similar, efficient algorithm for generating a subset of Q with probability proportional to its weight. Our empirical results suggest that block covers are superior to previous approaches, which consider the subsets of interest one by one.

Conformal prediction provides distribution-free coverage guarantees, but these degrade under distribution shift—and practitioners lack tools to anticipate which deployed models will fail before observing test data. We propose SHapley Additive exPlanations (SHAP) concentration—the fraction of feature importance concentrated in the top feature—as a pre-deployment diagnostic for conformal prediction vulnerability in gradient-boosted classifiers. Using COVID-19 as a naturalistic case study, eight supply chain tasks experience identical temporal shift yet coverage drops ranging from negligible to catastrophic. Feature-importance concentration is strongly associated with failure severity across 16 multiclass tasks in 9 domains, while standard distributional shift detectors detect shift uniformly across tasks but cannot distinguish catastrophic from robust outcomes. External validation across 9 non-supply-chain datasets shows partial transfer. We prove a formal theorem showing that Adaptive Prediction Sets conformity-score bounds worsen monotonically with concentration under explicit assumptions, verified empirically. The diagnostic identifies concentrated-dependence failures characteristic of gradient-boosted models but does not detect global-sensitivity failures observed in neural networks. A decision framework operationalizes the diagnostic as an exploratory pre-deployment rule with an uncertainty band around a concentration threshold.

Personalized text-to-image generation faces a fundamental trade-off between learning new concepts and preserving existing knowledge, and balancing these two goals gets particularly more challenging with multiple subjects to be personalized. Existing methods have relieved this by providing additional layout constraints, but it substantially undermines flexibility of the generation process. To address this issue, we propose a teacher–student architecture equipped with explicit regularizers to mitigate this trade-off. Specifically, we regulate the internal representations and cross-attention maps not to significantly deviate from the original foundation model, balanced with the reconstruction objective to internalize new concepts. Through extensive experiments, we demonstrate that our method enables reliable layout-free multi-subject generation, achieving the state-of-the-art performance on both single- and multi-concept personalization.