Human cognition can spontaneously shift conversation topics, often triggered by emotional or contextual signals. In contrast, self-attention-based language models depend on structured statistical cues from input tokens for next-token prediction, lacking this spontaneity. Motivated by this distinction, we investigate the factors that influence the next-token prediction to change the topic of the input sequence. We define concepts of topic continuity, ambiguous sequences, and change of topic, based on defining a topic as a set of token priority graphs (TPGs). Using a simplified single-layer self-attention architecture, we derive analytical characterizations of topic changes. Specifically, we demonstrate that (1) the model maintains the priority order of tokens related to the input topic, (2) a topic change occurs only if lower-priority tokens outnumber all higher-priority tokens of the input topic, and (3) unlike human cognition, longer context lengths and overlapping topics reduce the likelihood of spontaneous redirection. These insights highlight differences between human cognition and self-attention-based models in navigating topic changes and underscore the challenges in designing conversational AI capable of handling "spontaneous" conversations more naturally. To our knowledge, this is the first work to address these questions in such close relation to human conversation and thought.

The energy transition is profoundly reshaping electricity market dynamics. It makes it essential to understand how renewable energy generation actually impacts electricity prices, among all other market drivers. These insights are critical to design policies and market interventions that ensure affordable, reliable, and sustainable energy systems. However, identifying causal effects from observational data is a major challenge, requiring innovative causal inference approaches that go beyond conventional regression analysis only. We build upon the state of the art by developing and applying a local partially linear double machine learning approach. Its application yields the first robust causal evidence on the distinct and non-linear effects of wind and solar power generation on UK wholesale electricity prices, revealing key insights that have eluded previous analyses. We find that, over 2018-2024, wind power generation has a U-shaped effect on prices: at low penetration levels, a 1 GWh increase in energy generation reduces prices by up to 7 GBP/MWh, but this effect gets close to none at mid-penetration levels (20-30%) before intensifying again. Solar power places substantial downward pressure on prices at very low penetration levels (up to 9 GBP/MWh per 1 GWh increase in energy generation), though its impact weakens quite rapidly. We also uncover a critical trend where the price-reducing effects of both wind and solar power have become more pronounced over time (from 2018 to 2024), highlighting their growing influence on electricity markets amid rising penetration. Our study provides both novel analysis approaches and actionable insights to guide policymakers in appraising the way renewables impact electricity markets.

Recent photorealistic Novel View Synthesis (NVS) advances have increasingly gained attention. However, these approaches remain constrained to small indoor scenes. While optimization-based NVS models have attempted to address this, generalizable feed-forward methods, offering significant advantages, remain underexplored. In this work, we train PixelNeRF, a feed-forward NVS model, on the large-scale UrbanScene3D dataset. We propose four training strategies to cluster and train on this dataset, highlighting that performance is hindered by limited view overlap. To address this, we introduce Aug3D, an augmentation technique that leverages reconstructed scenes using traditional Structure-from-Motion (SfM). Aug3D generates well-conditioned novel views through grid and semantic sampling to enhance feed-forward NVS model learning. Our experiments reveal that reducing the number of views per cluster from 20 to 10 improves PSNR by 10%, but the performance remains suboptimal. Aug3D further addresses this by combining the newly generated novel views with the original dataset, demonstrating its effectiveness in improving the model's ability to predict novel views.

Scaling language models to handle longer input sequences typically necessitates large key-value (KV) caches, resulting in substantial memory overhead during inference. In this paper, we propose Tensor Product Attention (TPA), a novel attention mechanism that uses tensor decompositions to represent queries, keys, and values compactly, significantly shrinking KV cache size at inference time. By factorizing these representations into contextual low-rank components (contextual factorization) and seamlessly integrating with RoPE, TPA achieves improved model quality alongside memory efficiency. Based on TPA, we introduce the Tensor ProducT ATTenTion Transformer (T6), a new model architecture for sequence modeling. Through extensive empirical evaluation of language modeling tasks, we demonstrate that T6 exceeds the performance of standard Transformer baselines including MHA, MQA, GQA, and MLA across various metrics, including perplexity and a range of renowned evaluation benchmarks. Notably, TPA's memory efficiency enables the processing of significantly longer sequences under fixed resource constraints, addressing a critical scalability challenge in modern language models. The code is available at https://github.com/tensorgi/T6.

Despite the wide use of $k$-Nearest Neighbors as classification models, their explainability properties remain poorly understood from a theoretical perspective. While nearest neighbors classifiers offer interpretability from a "data perspective", in which the classification of an input vector $\bar{x}$ is explained by identifying the vectors $\bar{v}_1, \ldots, \bar{v}_k$ in the training set that determine the classification of $\bar{x}$, we argue that such explanations can be impractical in high-dimensional applications, where each vector has hundreds or thousands of features and it is not clear what their relative importance is. Hence, we focus on understanding nearest neighbor classifications through a "feature perspective", in which the goal is to identify how the values of the features in $\bar{x}$ affect its classification. Concretely, we study abductive explanations such as "minimum sufficient reasons", which correspond to sets of features in $\bar{x}$ that are enough to guarantee its classification, and "counterfactual explanations" based on the minimum distance feature changes one would have to perform in $\bar{x}$ to change its classification. We present a detailed landscape of positive and negative complexity results for counterfactual and abductive explanations, distinguishing between discrete and continuous feature spaces, and considering the impact of the choice of distance function involved. Finally, we show that despite some negative complexity results, Integer Quadratic Programming and SAT solving allow for computing explanations in practice.

Causal inference has recently gained notable attention across various fields like biology, healthcare, and environmental science, especially within explainable artificial intelligence (xAI) systems, for uncovering the causal relationships among multiple variables and outcomes. Yet, it has not been fully recognized and deployed in the manufacturing systems. In this paper, we introduce an explainable, scalable, and flexible federated Bayesian learning framework, \texttt{xFBCI}, designed to explore causality through treatment effect estimation in distributed manufacturing systems. By leveraging federated Bayesian learning, we efficiently estimate posterior of local parameters to derive the propensity score for each client without accessing local private data. These scores are then used to estimate the treatment effect using propensity score matching (PSM). Through simulations on various datasets and a real-world Electrohydrodynamic (EHD) printing data, we demonstrate that our approach outperforms standard Bayesian causal inference methods and several state-of-the-art federated learning benchmarks.

Spatial proteomics technologies have transformed our understanding of complex tissue architectures by enabling simultaneous analysis of multiple molecular markers and their spatial organization. The high dimensionality of these data, varying marker combinations across experiments and heterogeneous study designs pose unique challenges for computational analysis. Here, we present Virtual Tissues (VirTues), a foundation model framework for biological tissues that operates across the molecular, cellular and tissue scale. VirTues introduces innovations in transformer architecture design, including a novel tokenization scheme that captures both spatial and marker dimensions, and attention mechanisms that scale to high-dimensional multiplex data while maintaining interpretability. Trained on diverse cancer and non-cancer tissue datasets, VirTues demonstrates strong generalization capabilities without task-specific fine-tuning, enabling cross-study analysis and novel marker integration. As a generalist model, VirTues outperforms existing approaches across clinical diagnostics, biological discovery and patient case retrieval tasks, while providing insights into tissue function and disease mechanisms.

Age prediction using brain imaging, such as MRIs, has achieved promising results, with several studies identifying the model's residual as a potential biomarker for chronic disease states. In this study, we developed a brain age predictive model using a dataset of 1,220 U.S. veterans (18--80 years) and convolutional neural networks (CNNs) trained on two-dimensional slices of axial T2-weighted fast spin-echo and T2-weighted fluid attenuated inversion recovery MRI images. The model, incorporating a degree-3 polynomial ensemble, achieved an $R^{2}$ of 0.816 on the testing set. Images were acquired at the level of the anterior commissure and the frontal horns of the lateral ventricles. Residual analysis was performed to assess its potential as a biomarker for five ICD-coded conditions: hypertension (HTN), diabetes mellitus (DM), mild traumatic brain injury (mTBI), illicit substance abuse/dependence (SAD), and alcohol abuse/dependence (AAD). Residuals grouped by the number of ICD-coded conditions demonstrated different trends that were statistically significant ($p = 0.002$), suggesting a relationship between disease states and predicted brain age. This association was particularly pronounced in patients over 49 years, where negative residuals (indicating advanced brain aging) correlated with the presence of multiple ICD codes. These findings support the potential of residuals as biomarkers for detecting latent health conditions.

In Self-Supervised Learning (SSL), pre-training and evaluation are resource intensive. In the speech domain, current indicators of the quality of SSL models during pre-training, such as the loss, do not correlate well with downstream performance. Consequently, it is often difficult to gauge the final downstream performance in a cost efficient manner during pre-training. In this work, we propose unsupervised efficient methods that give insights into the quality of the pre-training of SSL speech models, namely, measuring the cluster quality and rank of the embeddings of the SSL model. Results show that measures of cluster quality and rank correlate better with downstream performance than the pre-training loss with only one hour of unlabeled audio, reducing the need for GPU hours and labeled data in SSL model evaluation.

Decision trees are widely used for classification and regression tasks in a variety of application fields due to their interpretability and good accuracy. During the past decade, growing attention has been devoted to globally optimized decision trees with deterministic or soft splitting rules at branch nodes, which are trained by optimizing the error function over all the tree parameters. In this work, we propose a new variant of soft multivariate regression trees (SRTs) where, for every input vector, the prediction is defined as the linear regression associated to a single leaf node, namely, the leaf node obtained by routing the input vector from the root along the branches with higher probability. SRTs exhibit the conditional computational property, i.e., each prediction depends on a small number of nodes (parameters), and our nonlinear optimization formulation for training them is amenable to decomposition. After showing a universal approximation result for SRTs, we present a decomposition training algorithm including a clustering-based initialization procedure and a heuristic for reassigning the input vectors along the tree. Under mild assumptions, we establish asymptotic convergence guarantees. Experiments on 15 wellknown datasets indicate that our SRTs and decomposition algorithm yield higher accuracy and robustness compared with traditional soft regression trees trained using the nonlinear optimization formulation of Blanquero et al., and a significant reduction in training times as well as a slightly better average accuracy compared with the mixed-integer optimization approach of Bertsimas and Dunn. We also report a comparison with the Random Forest ensemble method.