UAI 2026 Papers

We study identifiability of stochastic differential equations (SDE) under multiple interventions. Our results give the first provable bounds for unique recovery of SDE parameters given samples from their stationary distributions. We give tight bounds on the number of necessary interventions for linear SDEs, and upper bounds for nonlinear SDEs in the small noise regime. We experimentally validate the recovery of true parameters in synthetic data, and motivated by our theoretical results, demonstrate the advantage of parameterizations with learnable activation functions in application to gene regulatory dynamics.

Deep Q-Networks (DQNs) learn value functions through bootstrapped temporal-difference updates, where future returns are approximated using a greedy maximization over next-state action values. While effective, this aggregation rule is inherently sensitive to estimation noise: when Q-values are uncertain, the maximization operator deterministically favors the largest estimate, regardless of its reliability, leading to amplified errors through bootstrapping. In this work, we propose the **S**uccessor Rollout **A**ggregation **D**eep **Q**-Network (SADQ), a simple modification to Q-learning that regularizes how the TD target is formed. SADQ uses one-step rollout predictions from a learned dynamics model to guide the comparison among candidate next-state actions, introducing additional structure into the aggregation step without altering the underlying learning framework. The resulting mixed Bellman update attenuates unreliable maxima while preserving the standard fixed point under diminishing model error. We provide theoretical analysis showing that SADQ reduces bootstrap-induced overestimation in a pointwise manner. Empirically, SADQ consistently improves training stability across classical control tasks, real-world vector-based environments, and Atari benchmarks when compared to strong DQN variants.

In domains where statistical data is collected across hierarchically organized categories, drawing valid conclusions requires reasoning jointly about proportions and the structure of the domain. Statistical ℰℒ (𝒮ℰℒ) formalizes this kind of reasoning, but exact inference is ExpTime-hard and no implementation exists. We show how knowledge graph embeddings can approximate 𝒮ℰℒ inference efficiently. We prove analytical runtime and soundness guarantees, and empirically evaluate the runtime and approximation quality of our approach.

Cross-modal retrieval models often suffer substantial performance degradation under test-time distribution shifts. Existing test-time adaptation methods primarily based on entropy minimization tend to sharpen the similarity distribution. However, retrieval relies on feature similarities to rank candidates, and the similarity gaps between nearby-ranked candidates are often small. Over-sharpening the similarity gap between the candidates can distort the semantic similarity structure, leading to miscalibrated updates and unstable retrieval. To address this issue, we propose a novel *Unified Confidence Adjustment* framework that explicitly accounts for semantic similarity within the candidate set. Specifically, we propose the *Semantic Proximity Margin* as a confidence prior. Building on this margin, we develop a *Confidence State Identification* mechanism and a *Unified Confidence Adjustment* strategy for adaptive confidence calibration. Extensive experiments on four benchmarks under both zero-shot transfer and natural corruption settings show that our proposal consistently outperforms state-of-the-art test-time adaptation methods with minimal computational overhead, underscoring the effectiveness of confidence-aware regularization for robust cross-modal retrieval.

As deep learning models are increasingly deployed in high-stakes applications, providing well-calibrated uncertainty estimates has become as critical as achieving high predictive accuracy. While Kernel Density Estimation (KDE) has emerged as a smooth and continuous alternative to traditional binning for quantifying miscalibration, its reliability is heavily dependent on the choice of the kernel bandwidth. Standard selection techniques, such as Maximum Likelihood Estimation (MLE), often fail to produce optimal bandwidths for calibration tasks. In this work, we introduce Risk Alignment (RA), a novel optimization framework that determines the optimal bandwidth by aligning KDE-reconstructed risk with empirical risk. We theoretically demonstrate that this alignment minimizes calibration estimation bias across the data distribution, establishing a principled bandwidth selection criterion applicable to various metrics, including the challenging case of canonical calibration error. Extensive experiments across multiple architectures and datasets show that RA consistently outperforms standard bandwidth selection methods, yielding more reliable calibration assessments.

Modern e-commerce platforms employ various auction mechanisms to allocate paid slots for a given item. To scale this approach to the millions of auctions, the platforms suggest promotion tools based on the autobidding algorithms. These algorithms typically depend on the Click-Through-Rate (CTR) and Conversion-Rate (CVR) estimates provided by a pre-trained machine learning model. However, the predictions of such models are uncertain and can significantly affect the performance of the autobidding algorithm. To address this issue, we propose the method, which corrects the generated CTRs and CVRs to make the resulting bids more efficient in auctions. The underlying idea of our method is to employ a Bayesian approach and replace noisy CTR or CVR estimates with those from recovered distributions. To demonstrate the performance of the proposed approach, we perform extensive experiments on the synthetic, iPinYou, and BAT datasets. To evaluate the robustness of our approach to the noise scale, we use synthetic noise and noise estimated from the predictions of the pre-trained machine learning model.

Super-resolution (SR) of satellite imagery is challenging due to the lack of paired low-/high-resolution data. Recent self-supervised SR methods overcome this limitation by exploiting the temporal redundancy in burst observations, but they lack a mechanism to quantify uncertainty in the reconstruction. In this work, we introduce a novel self-supervised loss that allows to estimate uncertainty in image super-resolution without ever accessing the ground-truth high-resolution data. We adopt a decision-theoretic perspective and show that minimizing the corresponding Bayesian risk yields the posterior mean and variance as optimal estimators. We validate our approach on a synthetic dataset with both white and signal dependent noise and demonstrate that it produces calibrated uncertainty estimates comparable to supervised methods. We further apply our method on real SkySat satellite data and validate its performance through an unsupervised coverage test. Our work bridges self-supervised restoration with uncertainty quantification, making a practical framework for uncertainty-aware image reconstruction.

Reliable uncertainty quantification (UQ) is essential for deploying large language models (LLMs) in safety-critical scenarios, as it enables them to abstain from responding when uncertain, thereby avoiding hallucinations, i.e., plausible yet factually incorrect responses. However, while current semantic UQ methods have achieved state-of-the-art performance, they inherently overlook latent semantic structural information that could enable more precise uncertainty estimates. In this paper, we propose Semantic Structural Entropy (SeSE), a principled black-box UQ framework applicable to both open- and closed-source LLMs. To reveal the intrinsic structure of the LLM semantic space, SeSE constructs its hierarchical abstraction based on the principle of structural entropy minimization. The structural entropy of the resulting optimal hierarchical abstraction thus quantifies the inherent uncertainty within the semantic space after optimal compression. Additionally, unlike existing methods that primarily focus on simple short-form generation, we extend SeSE to provide interpretable and granular uncertainty estimation for long-form outputs. We theoretically prove that SeSE generalizes semantic entropy, the gold standard for UQ in LLMs, and empirically demonstrate its superior performance over baselines across 24 model-dataset combinations.

We study a solution concept for bounded rational agents in finite normal-form general-sum games called Generalized Quantal Response Equilibrium (GQRE) which generalizes a Quantal Response Equilibrium (McKelvey and Palfrey, 1995). In our setup, each player can individually maximize a smooth, regularized expected utility of the mixed profiles used, reflecting both bounded rationality that subsumes stochastic choice, and also individual choice of behaviors. After establishing existence under mild conditions, we present a computationally efficient no-regret decentralized learning algorithm that uses a smoothened version of the Frank–Wolfe algorithm coupled with a computationally efficient projection step. Our algorithm uses noisy gradient estimates via bandit-feedback from a simulation oracle that reports on repeated plays of the game. We analyze finite-time convergence properties of our algorithm under assumptions that ensure uniqueness of equilibrium, using a novel class of gap functions that generalize the Nash gap function. We end by demonstrating the effectiveness of our method on a set of complex general-sum games such as high-rank two-player games, large action two-player games, and known examples of difficult multi-player games.

Diffusion Models (DMs) have achieved remarkable success in image generation, yet recent studies reveal their vulnerability to backdoor attacks, where adversaries manipulate outputs via covert triggers embedded in inputs. Existing defenses, such as backdoor detection and trigger inversion, are largely effective because prior attacks rely on limited input spaces and low-dimensional triggers that are visually conspicuous or easily captured by neural detectors. To broaden the threat landscape, we propose Gungnir, a novel backdoor attack that activates malicious behaviors through style-based triggers embedded in input images. Unlike explicit visual patches or textual cues, stylistic features serve as stealthy, high-level triggers. We introduce Reconstructing-Adversarial Noise (RAN) and Short-Term Timesteps-Retention (STTR) to preserve trigger-consistent diffusion dynamics in image-to-image tasks. The resulting trigger-embedded samples are perceptually indistinguishable from clean images, evading both manual and automated detection. Extensive experiments show that Gungnir bypasses state-of-the-art defenses with an extremely low backdoor detection rate (BDR) and remains effective under fine-tuning-based purification, revealing previously underexplored vulnerabilities in diffusion models.

Instrumental variable (IV) and proximal causal learning (Proxy) methods are central frameworks for causal inference in the presence of unobserved confounding. Despite substantial methodological advances, existing approaches rarely provide reliable epistemic uncertainty (EU) quantification. We address this gap through a Deconditional Gaussian Process (GP) framework for uncertainty-aware causal learning. Our formulation recovers popular kernel estimators as the posterior mean, ensuring predictive precision, while the posterior variance yields principled and well-calibrated EU. Moreover, the probabilistic structure enables systematic model selection via marginal log-likelihood optimization. Empirical results demonstrate strong predictive performance alongside informative EU quantification, evaluated via empirical coverage frequencies and decision-aware accuracy–rejection curves. Together, our approach provides a unified, practical solution for causal inference under unobserved confounding with reliable uncertainty.

Scaling test-time computation enhances LLM reasoning ability but faces a uniform computation paradox. Allocating identical resources leads to over-correction on simple tasks and insufficient refinement on complex ones. To address this, we propose CoFiCot, a coarse-to-fine adaptive framework that dynamically tailors inference strategies to problem difficulty. Specifically, we implement a multi-metric classifier that triages queries by synthesizing semantic entropy, consensus reliability, and predicted reasoning depth . This enables a differentiated refinement stage that applies efficient aggregation for simple queries while routing complex ones to a context-aware correction loop . We formalize correction as a stateful sequential propagation process , where each repair is strictly conditioned on the verified history of prior rectifications. By integrating Process Reward Models (PRMs) within this state-dependent trajectory, CoFiCot effectively bridges the gap between granular error localization and global logical coherence, preventing the context fragmentation typical of stateless refinement methods.

Federated learning (FL) is often subject to aggregation variance if clients do not consistently participate in training rounds. While reusing stale model updates from inactive clients is a common technique to reduce this variance, we find that with skewed client participation, the resulting update staleness can become severe enough to destabilize training. To remedy this, we propose FedSteer, a novel method that constructs a gradient subspace from a cache of recent client gradients to serve as a low-dimensional representation of the current optimization landscape. FedSteer projects an active client’s true gradient onto this subspace to find a set of optimal coordinates. For an inactive client, FedSteer reuses these coordinates with the now-evolved subspace drifted by other active clients. This process effectively “steers” outdated gradients toward the current global objective. This is complemented by a selective caching strategy that identifies a representative client subset to form the subspace, reducing server memory. Experiments demonstrate that FedSteer significantly outperforms baselines, preventing performance collapse in challenging scenarios while delivering accuracy gains of over 7

Causal discovery aims to uncover the underlying causal relationships given data generated from a system. The goal, however, is not merely to predict causal edges given data, but also to be able to interpret and explain either observed or hypothesized phenomena, such as a particularly large causal effect. We consider this task of conditional causal discovery and cast it as a Bayesian inference problem, in which we target the posterior over causal graphs and parameters conditional on an event such as a causal-effect constraint. Unfortunately, this poses a computational challenge: existing approaches to Bayesian causal discovery struggle when the event has small posterior mass. To address this, we adapt rare-event estimation techniques to perform inference the joint graph-parameter space. Our method gradually drives a particle population toward the constrained region while maintaining samples that approximate the conditional posterior. Empirical evaluation on synthetic graphs validates the accuracy of our approach at small and large scales, and we show in a case study on the Sachs protein dataset how our method can be used to aid scientific exploration by providing pathway-level summaries.

Predictive machine learning (ML) models are increasingly used to aid human decision-makers across various high-risk domains such as healthcare and criminal justice. There is a growing recognition of the need to evaluate the causal impact of deploying these systems on downstream outcomes, such as patient survival or crime recidivism. Randomized control trials (RCTs) can provide high-quality evidence on the impact of a deployed model, but they run into a challenge: it is often infeasible to run repeated trials when models are updated or retrained to improve predictive performance. In this work, we present a partial-identification approach to using prior RCT data to construct bounds on the causal effect of a new model. The core innovation in our approach is to leverage assumptions relating fine-grained predictive accuracy to downstream outcomes. We do so via two monotonicity assumptions: first, on individual-level ‘counterfactual correctness’ (all else being equal, a correct prediction leads to non-inferior outcomes); and second, on the relation between subgroup predictive performance and outcomes, interpretable as an assumption regarding trust in model outputs. We demonstrate our method with a simulation study, illustrating how incorporating this information can lead to more informative bounds compared to prior work.

Intracortical brain-computer interfaces (BCIs) can decode speech from neural activity with high accuracy when trained on data pooled across recording sessions. In realistic deployment, however, models must generalize to new sessions without labeled data, and performance often degrades due to cross-session nonstationarities (e.g., electrode shifts, neural turnover, and changes in user strategy). In this paper, we propose ALIGN, a session-invariant learning framework based on multi-domain adversarial neural networks for semi-supervised cross-session adaptation. ALIGN trains a feature encoder jointly with a phoneme classifier and a domain classifier operating on the latent representation. Through adversarial optimization, the encoder is encouraged to preserve task-relevant information while suppressing session-specific cues. We evaluate ALIGN on intracortical speech decoding and find that it generalizes consistently better to previously unseen sessions, improving both phoneme error rate and word error rate relative to baselines. These results indicate that adversarial domain alignment is an effective approach for mitigating session-level distribution shift and enabling robust longitudinal BCI decoding.

Decision trees are widely used in machine learning due to their interpretability. However, when the underlying data distribution changes—a phenomenon known as concept drift—their performance can degrade significantly. Retraining the model from scratch discards the learned structure and offers no insight into which nodes were actually affected. Prior work, APPETITE, used spectrum-based fault localization to identify and modify a single faulty node, but cannot handle drift that affects multiple nodes simultaneously or that demands structural changes to the tree. We propose BTAD (Bayesian Tree Adaptation for Drift), which formalizes the Decision Tree Concept Drift Problem as a Model-Based Diagnosis task and employs BARINEL, a Bayesian algorithm that ranks multi-node fault hypotheses across the entire tree. For repair, BTAD applies Subtree Re-fitting, replacing affected subtrees with ones fitted to the post-drift distribution. BTAD relies on lightweight node-level statistics and requires no access to the original training data. Experiments show that BTAD outperforms APPETITE in both recovery and localization, achieving accuracy comparable to full retraining—which provides no drift insight and requires the original data.

The LASSO estimator has been popular in diverse fields such as statistics, bio-informatics and compressed sensing for estimation of sparse signals from under-sampled measurements. As the LASSO does not provide confidence intervals for the estimates of the individual elements, techniques such as the debiased LASSO have been developed that provide such confidence intervals and also mitigate the inherent bias in the LASSO estimates. However the theoretical guarantees underlying the debiased LASSO have been developed for a design or measurement matrix whose elements have zero mean. On the other hand, many commonly used measurement matrices in sparse regression applications in optical imaging or pooled biological testing involve non-negative, particularly binary, matrices whose elements have a non-zero mean. In this paper, we propose a debiased LASSO technique based on computing differences between different row subsets of the underlying design matrix iteratively which effectively converts the design matrix into one with zero-mean elements. We show analytically that this results in lower variances for the individual elements of the debiased LASSO estimate as compared to the standard debiased LASSO. We also provide numerical results to support our theory.

Uncertainty quantification is central to safe and efficient deployments of deep learning models, yet many computationally practical methods lack lacking rigorous theoretical motivation. Random network distillation (RND) is a lightweight technique that measures novelty via prediction errors against a fixed random target. While empirically effective, it has remained unclear what uncertainties RND measures and how its estimates relate to other approaches, e.g., Bayesian inference or deep ensembles. We establish these missing theoretical connections by analyzing RND within the neural tangent kernel framework in the limit of infinite network width. Our analysis reveals two central findings in this limit: (1) The uncertainty signal from RND—its squared self-predictive error—is equivalent to the predictive variance of a deep ensemble. (2) By constructing a specific RND target function, we show that the RND error distribution can be made to mirror the centered posterior predictive distribution of Bayesian inference with wide neural networks. Based on this equivalence, we moreover devise a posterior sampling algorithm that generates i.i.d. samples from an exact Bayesian posterior predictive distribution using this modified Bayesian RND model. Collectively, our findings provide a unified theoretical perspective that places RND within the principled frameworks of deep ensembles and Bayesian inference, and offer new avenues for efficient yet theoretically grounded uncertainty quantification methods.