The call for tutorials was here and the list of accepted (long and short) tutorials is below.
Peter Grünwald (CWI/Leiden University); Rianne de Heide (CWI/University of Twente)
We give a basic yet comprehensive introduction to e-values, a notion of statistical evidence
and uncertainty quantification (UQ) that is rapidly becoming more popular. We provide examples of practically useful e-values, indicate how to construct e-values for a given testing or
UQ setting, and explain their two most distinguishing uses: anytime-valid inference, and setting
significance/confidence level in post-hoc, data-dependent manners.
Yuejie Chi (Yale University); Harry Dong (Microsoft Research)
Large language models (LLMs) have achieved tremendous successes in artificial intelligence, where transformers with the attention mechanism serve as the backbone architecture. Chain-of-thought (CoT) significantly improves the reasoning capabilities of transformers on complex tasks like math and agents by generating intermediate reasoning steps, before producing the final answers. By eliciting this behavior via supervised fine-tuning (SFT) and reinforcement learning (RL), CoT leads to a paradigm shift to test-time scaling represented by the rise of reasoning models, which focus on improving the model performance via allocating more compute at the time of inference. This tutorial will provide an introduction to CoT reasoning in transformers from both theoretical and empirical perspectives. We explore theoretical questions around training dynamics, model expressivity, length generalization and reasoning mechanisms, and practical considerations including efficiency, scalability, data mix, and parallel scaling. The tutorial aims to bridge emerging theoretical insights with practical advances in understanding, building, and deploying reasoning-capable LLMs.
Giuseppe Marra (KU Leuven); Pietro Barbiero (IBM Research); David Debot (KU Leuven)
Probabilistic Concept Bottleneck Models (CBMs) provide a principled framework for building interpretable and structured machine learning (ML) systems by explicitly modeling high-level concepts and their uncertainty. This tutorial introduces participants to the foundations, methodologies, and practical implementations of probabilistic CBMs, positioning them within broader themes of interpretability, explainability, and alignment. We begin by motivating concept-based approaches as an alternative to post-hoc explanations, emphasizing their advantages for transparency and human interaction. We then present probabilistic extensions of classical CBMs, highlighting how uncertainty modeling enables more robust reasoning and effective human-in-the-loop interventions. The tutorial further surveys recent advances, including causal and stochastic CBMs, as well as neurosymbolic (NeSy) approaches that integrate structured reasoning with deep learning (DL). Finally, participants will gain hands-on experience by implementing CBMs using PyC, a modern PyTorch-based library for concept-based modeling. By the end of the tutorial, attendees will understand the theoretical foundations of probabilistic CBMs, recognize their connections to related fields such as NeSy, and acquire practical skills to apply these models in real-world settings.
Willem Zuidema (University of Amsterdam); Ana Lucic (University of Amsterdam)
Interpretability methods are a class of analysis methods aimed at revealing how/why modern machine learning models make particular predictions. As machine learning permeates science and society, these methods –and a good understanding of their appropriate uses and limitations– have become ever more important. This tutorial provides a systematic overview of the current landscape of explainability/interpretability, with a particular emphasis on recent advances that (1) go beyond traditional attribution-based approaches and (2) are applied to the natural sciences, where they promise to help establish interfaces with existing theories and concepts. We first cover the theoretical foundations and the practical considerations/limitations of the main families of XAI methods, including feature attribution, concept-based explanations, and mechanistic interpretability. Next, we cover techniques that are specific to different types of architectures that are now dominant in (scientific) ML: transformers, graph neural networks, time series, and diffusion models. Finally, we examine how these techniques are used in the context of real-world applications, specifically in the natural sciences: neuroscience, weather/climate, and molecular simulation.
Junhyung Park (ETH Zürich)
Causal reasoning is usually formalised through structural causal models (SCMs) or potential outcomes. These frameworks have been enormously successful for modelling, identification, and inference, but they are not primarily designed as axiomatic foundations analogous
to probability spaces in probability theory. This tutorial introduces causal spaces, a measure-theoretic framework in which interventions are represented by primitive causal kernels satisfying two minimal axioms: doing nothing changes nothing, and intervened coordinates take
their prescribed values. The tutorial will explain the motivation for causal spaces, present the basic definition and semantics, and work through examples linking back to familiar causal models. It will then present some further development of basic causal space theory, such as causal effects, sources, identifiability and counterfactual spaces, and touch upon advanced topics such as targeted interventions and continuous-time stochastic processes that are more difficult to express in existing frameworks. The goal is to make causal spaces accessible to the broader UAI audience, while giving theoretically oriented participants enough detail to engage with current research directions.
Morten Mørup (Technical University of Denmark); Kazu Ghalamkari (Technical University of Denmark)
How can we estimate the true distribution underlying the given data? This is one of the fundamental questions in machine learning. With the current advances in GPU accelerators, the community relies heavily on deep learning-based density estimation, which offers great success in expressivity and scalability; however, theoretical intractability, the need for costly hyper-parameter tuning, weaker optimization guarantees, and the non-trivial extension to the discrete settings remain important challenges. Recently, alternative approaches for density estimation, centered on the use of tensor networks, are garnering attention as they can overcome these difficulties. In addition, their connections have also recently been established to other fields such as probabilistic circuits, information geometry, logic programming, and relational learning, forming a rich community in which tensors play a role of shared language, as seen in recent tensor-related workshops, Connecting Low-Rank Representations in AI at AAAI’25 and ICML’26, and a tutorial, Foundations of Tensor/Low-Rank Computations for AI at Neurips 2025. Given the current situation, we presently provide a tutorial on tensor-based density estimation where its exact marginalization, natural Bayesian extension, and convergence guarantees directly match the interests of the UAI community. Aiming to welcome newcomers as well as bridging various fields, this tutorial covers the following topics: i) How tensors are useful for density estimation, ii) how tensor-based density estimation connects diverse fields, and iii) what are important future directions of tensor networks for the UAI community.
Cassio de Campos (TU Eindhoven)
Credal networks are probabilistic graphical models based on imprecise probabilities. They can be regarded as an extension of Bayesian networks, where credal sets replace probability distributions in the specification of the local models for the network variables given their parents. While a Bayesian network defines a joint probability distribution over its variables, a credal network defines a joint credal set of probability distributions. This allows us greater flexibility when modelling problems and also creates connections with other learning and reasoning tasks, including (marginal) MAP inferences, sensitivity analyses, robust and adversarial situations. This tutorial will discuss these models and their relations with other models and inferential tasks. In particular, we will connect with probabilistic circuits and Markov networks. Finally, we will discuss credal approaches as means of creating privacy-aware models and in the scope of reliable classification and conformal predictions.
Eleonora Giunchiglia (Imperial College London); Mihaela Catalina Stoian (Imperial College London)
Many applications of machine learning require neural networks to produce outputs that satisfy hard constraints. Examples include portfolio weights that must obey regulatory diversification rules, generated tabular records that must satisfy domain-specific consistency requirements, and continuous decisions that must remain inside safety-critical feasible regions. Yet standard deep learning methods typically handle such requirements through soft penalties, post-hoc projection, or rejection sampling, none of which provides a general and reliable mechanism for enforcing complex constraints at prediction time. This tutorial will provide a self-contained introduction to methods for constraining neural networks whose outputs are continuous variables. The tutorial will first introduce the mathematical foundations of constrained neural prediction, including the distinction between soft regularisation, exact enforcement, differentiable correction layers, and optimisation-based neural layers. It will then present a set of representative state-of-the-art approaches, including Deep Constraint Completion and Correction (DC3), the Disjunctive Refinement Layer (DRL), LinSATNet, and the Probabilistic Algebraic Layer (PAL). For each method, we will discuss the constraint language it supports, the type of guarantees it provides, its computational trade-offs, and its suitability for integration into modern deep learning pipelines. By the end of the tutorial, attendees will understand how to move beyond soft-penalty heuristics and select, implement, and critically evaluate neural architectures that enforce hard constraints over continuous output spaces.
Last updated: May 27, 2026 16:02 (UTC)
Sponsors: TBD