Graph Neural Networks (GNNs) have proven to be highly effective in various graph learning tasks. A key characteristic of GNNs is their use of a fixed number of message-passing steps for all nodes in the graph, regardless of each node's diverse computational needs and characteristics. Through empirical real-world data analysis, we demonstrate that the optimal number of message-passing layers varies for nodes with different characteristics. This finding is further supported by experiments conducted on synthetic datasets. To address this, we propose Adaptive Depth Message Passing GNN (ADMP-GNN), a novel framework that dynamically adjusts the number of message passing layers for each node, resulting in improved performance. This approach applies to any model that follows the message passing scheme. We evaluate ADMP-GNN on the node classification task and observe performance improvements over baseline GNN models.

AdamW has long been the dominant optimizer in language model pretraining, despite numerous claims that alternative optimizers offer 1.4 to 2x speedup. We posit that two methodological shortcomings have obscured fair comparisons and hindered practical adoption: (i) unequal hyperparameter tuning and (ii) limited or misleading evaluation setups. To address these two issues, we conduct a systematic study of ten deep learning optimizers across four model scales (0.1B-1.2B parameters) and data-to-model ratios (1-8x the Chinchilla optimum). We find that fair and informative comparisons require rigorous hyperparameter tuning and evaluations across a range of model scales and data-to-model ratios, performed at the end of training. First, optimal hyperparameters for one optimizer may be suboptimal for another, making blind hyperparameter transfer unfair. Second, the actual speedup of many proposed optimizers over well-tuned baselines is lower than claimed and decreases with model size to only 1.1x for 1.2B parameter models. Thirdly, comparing intermediate checkpoints before reaching the target training budgets can be misleading, as rankings between two optimizers can flip during training due to learning rate decay. Through our thorough investigation, we find that all the fastest optimizers such as Muon and Soap, use matrices as preconditioners -- multiplying gradients with matrices rather than entry-wise scalars. However, the speedup of matrix-based optimizers is inversely proportional to model scale, decreasing from 1.4x over AdamW for 0.1B parameter models to merely 1.1x for 1.2B parameter models.

We study the problem of evaluating the excess risk of classical penalized empirical risk minimization (ERM) with Bregman losses. We show that by leveraging the recently proposed wild refitting procedure (Wainwright, 2025), one can efficiently upper bound the excess risk through the so-called "wild optimism," without relying on the global structure of the underlying function class. This property makes our approach inherently model-free. Unlike conventional analyses, our framework operates with just one dataset and black-box access to the training procedure. The method involves randomized vector-valued symmetrization with an appropriate scaling of the prediction residues and constructing artificially modified outcomes, upon which we retrain a second predictor for excess risk estimation. We establish high-probability performance guarantees both under the fixed design setting and the random design setting, demonstrating that wild refitting under Bregman losses, with an appropriately chosen wild noise scale, yields a valid upper bound on the excess risk. This work thus is promising for theoretically evaluating modern opaque ML and AI models such as deep neural networks and large language models, where the model class is too complex for classical learning theory and empirical process techniques to apply.

Operator learning has emerged as a promising tool for accelerating the solution of partial differential equations (PDEs). The Deep Operator Networks (DeepONets) represent a pioneering framework in this area: the "vanilla" DeepONet is valued for its simplicity and efficiency, while the modified DeepONet achieves higher accuracy at the cost of increased training time. In this work, we propose a series of Transformer-inspired DeepONet variants that introduce bidirectional cross-conditioning between the branch and trunk networks in DeepONet. Query-point information is injected into the branch network and input-function information into the trunk network, enabling dynamic dependencies while preserving the simplicity and efficiency of the "vanilla" DeepONet in a non-intrusive manner. Experiments on four PDE benchmarks -- advection, diffusion-reaction, Burgers', and Korteweg-de Vries equations -- show that for each case, there exists a variant that matches or surpasses the accuracy of the modified DeepONet while offering improved training efficiency. Moreover, the best-performing variant for each equation aligns naturally with the equation's underlying characteristics, suggesting that the effectiveness of cross-conditioning depends on the characteristics of the equation and its underlying physics. To ensure robustness, we validate the effectiveness of our variants through a range of rigorous statistical analyses, among them the Wilcoxon Two One-Sided Test, Glass's Delta, and Spearman's rank correlation.

This paper investigates the expected excess risk of In-Context Learning (ICL) for multi-class classification. We model each task as a sequence of labeled prompt samples and a query input, where a pre-trained model estimates the conditional class probabilities of the query. The expected excess risk is defined as the average truncated Kullback-Leibler (KL) divergence between the predicted and ground-truth conditional class distributions, averaged over a specified family of tasks. We establish a new oracle inequality for the expected excess risk based on KL divergence in multiclass classification. This allows us to derive tight upper and lower bounds for the expected excess risk in transformer-based models, demonstrating that the ICL estimator achieves the minimax optimal rate--up to a logarithmic factor--for conditional probability estimation. From a technical standpoint, our results introduce a novel method for controlling generalization error using the uniform empirical covering entropy of the log-likelihood function class. Furthermore, we show that multilayer perceptrons (MLPs) can also perform ICL and achieve this optimal rate under specific assumptions, suggesting that transformers may not be the exclusive architecture capable of effective ICL.

At its core, machine learning seeks to train models that reliably generalize beyond noisy observations; however, the theoretical vacuum in which state-of-the-art universal approximation theorems (UATs) operate isolates them from this goal, as they assume noiseless data and allow network parameters to be chosen freely, independent of algorithmic realism. This paper bridges that gap by introducing an architecture-specific randomized training algorithm that constructs a uniform approximator from $N$ noisy training samples on the $d$-dimensional cube $[0,1]^d$. Our trained neural networks attain the minimax-optimal quantity of \textit{trainable} (non-random) parameters, subject to logarithmic factors which vanish under the idealized noiseless sampling assumed in classical UATs. Additionally, our trained models replicate key behaviours of real-world neural networks, absent in standard UAT constructions, by: (1) exhibiting sub-linear parametric complexity when fine-tuning on structurally related and favourable out-of-distribution tasks, (2) exactly interpolating the training data, and (3) maintaining reasonable Lipschitz regularity (after the initial clustering attention layer). These properties bring state-of-the-art UATs closer to practical machine learning, shifting the central open question from algorithmic implementability with noisy samples to whether stochastic gradient descent can achieve comparable guarantees.

While electroencephalogram (EEG) has been a crucial tool for monitoring the brain and diagnosing neurological disorders (e.g., epilepsy), learning meaningful representations from raw EEG signals remains challenging due to limited annotations and high signal variability. Recently, EEG foundation models (FMs) have shown promising potential by adopting transformer architectures and self-supervised pre-training methods from large language models (e.g., masked prediction) to learn representations from diverse EEG data, followed by fine-tuning on specific EEG tasks. Nonetheless, these large models often incurred high computational costs during both training and inference, with only marginal performance improvements as the model size increases. In this work, we proposed an EEG representation learning framework building upon Generative Diffusion Model (EEGDM). Specifically, we developed a structured state-space model for diffusion pretraining (SSMDP) to better capture the temporal dynamics of EEG signals and trained it using Denoising Diffusion Probabilistic Model (DDPM) framework. Subsequently, the resulting latent EEG representations were then used for downstream classification tasks via our proposed latent fusion transformer (LFT). To evaluate our method, we used multi-event datasets covering both interictal epileptiform discharges (TUEV) and seizure (CHB-MIT) detection, and compared EEGDM with current state-of-the-art approaches, including EEG FMs. Empirical results showed that our method outperformed the existing methods. These findings suggested that EEGDM offered a promising alternative to current FMs. Our source code and checkpoint are available at: https://github.com/jhpuah/EEGDM.

Mental disorders are clinically significant patterns of behavior that are associated with stress and/or impairment in social, occupational, or family activities. People suffering from such disorders are often misjudged and poorly diagnosed due to a lack of visible symptoms compared to other health complications. During emergency situations, identifying psychiatric issues is that's why challenging but highly required to save patients. In this paper, we have conducted research on how traditional machine learning and large language models (LLM) can assess these psychiatric patients based on their behavioral patterns to provide a diagnostic assessment. Data from emergency psychiatric patients were collected from a rescue station in Germany. Various machine learning models, including Llama 3.1, were used with rescue patient data to assess if the predictive capabilities of the models can serve as an efficient tool for identifying patients with unhealthy mental disorders, especially in rescue cases.

Transformer-based models such as BERT and E5 have significantly advanced text embedding by capturing rich contextual representations. However, many complex real-world queries require sophisticated reasoning to retrieve relevant documents beyond surface-level lexical matching, where encoder-only retrievers often fall short. Decoder-only large language models (LLMs), known for their strong reasoning capabilities, offer a promising alternative. Despite this potential, existing LLM-based embedding methods primarily focus on contextual representation and do not fully exploit the reasoning strength of LLMs. To bridge this gap, we propose Reasoning-Infused Text Embedding (RITE), a simple but effective approach that integrates logical reasoning into the text embedding process using generative LLMs. RITE builds upon existing language model embedding techniques by generating intermediate reasoning texts in the token space before computing embeddings, thereby enriching representations with inferential depth. Experimental results on BRIGHT, a reasoning-intensive retrieval benchmark, demonstrate that RITE significantly enhances zero-shot retrieval performance across diverse domains, underscoring the effectiveness of incorporating reasoning into the embedding process.

Recently, large language models (LLMs) have achieved widespread application across various fields. Despite their impressive capabilities, LLMs suffer from a lack of structured reasoning ability, particularly for complex tasks requiring domain-specific best practices, which are often unavailable in the training data. Although multi-step prompting methods incorporating human best practices, such as chain-of-thought and tree-of-thought, have gained popularity, they lack a general mechanism to control LLM behavior. In this paper, we propose SHERPA, a model-driven framework to improve the LLM performance on complex tasks by explicitly incorporating domain-specific best practices into hierarchical state machines. By structuring the LLM execution processes using state machines, SHERPA enables more fine-grained control over their behavior via rules or decisions driven by machine learning-based approaches, including LLMs. We show that SHERPA is applicable to a wide variety of tasks-specifically, code generation, class name generation, and question answering-replicating previously proposed approaches while further improving the performance. We demonstrate the effectiveness of SHERPA for the aforementioned tasks using various LLMs. Our systematic evaluation compares different state machine configurations against baseline approaches without state machines. Results show that integrating well-designed state machines significantly improves the quality of LLM outputs, and is particularly beneficial for complex tasks with well-established human best practices but lacking data used for training LLMs.