Fresh from the feed

arXiv:2509.17866v2 Announce Type: replace Abstract: Post-training fundamentally alters the behavior of large language models (LLMs), yet its impact on the internal parameter space remains poorly understood. In this work, we conduct a systematic singular value decomposition (SVD) analysis of principal linear layers in pretrained LLMs, focusing on two widely adopted post-training methods: instruction tuning and long-chain-of-thought (Long-CoT) distillation. Our analysis reveals two consistent and unexpected structural changes:(1) a near-uniform geometric scaling of singular values across layers, which theoretically modulates attention scores; and (2) highly consistent orthogonal transformations are applied to the left and right singular vectors of each matrix. Disrupting this orthogonal consistency leads to catastrophic performance degradation. Based on these findings, we propose a simple yet effective framework that interprets post-training as a reparameterization of fixed subspaces in the pretrained parameter space. Further experiments reveal that singular value scaling behaves as a secondary effect, analogous to a temperature adjustment, whereas the core functional transformation lies in the coordinated rotation of singular vectors. These results challenge the prevailing view of the parameter space in large models as a black box, uncovering the first clear regularities in how parameters evolve during training, and providing a new perspective for deeper investigation into model parameter changes.

arXiv:2509.17208v2 Announce Type: replace Abstract: Machine-learned coarse-grained (CG) potentials are fast, but degrade over time when simulations reach under-sampled bio-molecular conformations, and generating widespread all-atom (AA) data to combat this is computationally infeasible. We propose a novel active learning (AL) framework for CG neural network potentials in molecular dynamics (MD). Building on the CGSchNet model, our method employs root mean squared deviation (RMSD)-based frame selection from MD simulations in order to generate data on-the-fly by querying an oracle during the training of a neural network potential. This framework preserves CG-level efficiency while correcting the model at precise, RMSD-identified coverage gaps. By training CGSchNet, a coarse-grained neural network potential, we empirically show that our framework explores previously unseen configurations and trains the model on unexplored regions of conformational space. Our active learning framework enables a CGSchNet model trained on the Chignolin protein to achieve a 33.05\% improvement in the Wasserstein-1 (W1) metric in Time-lagged Independent Component Analysis (TICA) space on an in-house benchmark suite.

arXiv:2509.16625v3 Announce Type: replace Abstract: Detecting intrusions in network traffic is a challenging task, particularly under limited supervision and constantly evolving attack patterns. While recent works have leveraged graph neural networks for network intrusion detection, they often decouple representation learning from anomaly detection, limiting the utility of the embeddings for identifying attacks. We propose GraphIDS, a self-supervised intrusion detection model that unifies these two stages by learning local graph representations of normal communication patterns through a masked autoencoder. An inductive graph neural network embeds each flow with its local topological context to capture typical network behavior, while a Transformer-based encoder-decoder reconstructs these embeddings, implicitly learning global co-occurrence patterns via self-attention without requiring explicit positional information. During inference, flows with unusually high reconstruction errors are flagged as potential intrusions. This end-to-end framework ensures that embeddings are directly optimized for the downstream task, facilitating the recognition of malicious traffic. On diverse NetFlow benchmarks, GraphIDS achieves up to 99.98% PR-AUC and 99.61% macro F1-score, outperforming baselines by 5-25 percentage points.

arXiv:2509.16339v3 Announce Type: replace Abstract: We investigate imbalanced regression with tabular data that have an imbalance ratio larger than 1,000 ("highly imbalanced"). Accurately estimating the target values of rare instances is important in applications such as forecasting the intensity of rare harmful Solar Energetic Particle (SEP) events. For regression, the MSE loss does not consider the correlation between predicted and actual values. Typical inverse importance functions allow only convex functions. Uniform sampling might yield mini-batches that do not have rare instances. We propose CISIR that incorporates correlation, Monotonically Decreasing Involution (MDI) importance, and stratified sampling. Based on five datasets, our experimental results indicate that CISIR can achieve lower error and higher correlation than some recent methods. Also, adding our correlation component to other recent methods can improve their performance. Lastly, MDI importance can outperform other importance functions. Our code can be found in https://github.com/Machine-Earning/CISIR.

arXiv:2509.10517v2 Announce Type: replace Abstract: Machine learning models hold significant potential for predicting in-hospital mortality, yet data privacy constraints and the statistical heterogeneity of real-world clinical data often hamper their development. Federated Learning (FL) offers a privacy-preserving solution, but its performance under non-Independent and Identically Distributed (non-IID) and imbalanced conditions requires rigorous investigation. The study presents a comparative benchmark of five federated learning strategies: FedAvg, FedProx, FedAdagrad, FedAdam, and FedCluster for mortality prediction. Using the large-scale MIMIC-IV dataset, we simulate a realistic non-IID environment by partitioning data by clinical care unit. To address the inherent class imbalance of the task, the SMOTE-Tomek technique is applied to each client's local training data. Our experiments, conducted over 50 communication rounds, reveal that the regularization-based strategy, FedProx, consistently outperformed other methods, achieving the highest F1-Score of 0.8831 while maintaining stable convergence. While the baseline FedAvg was the most computationally efficient, its predictive performance was substantially lower. Our findings indicate that regularization-based FL algorithms like FedProx offer a more robust and effective solution for heterogeneous and imbalanced clinical prediction tasks than standard or server-side adaptive aggregation methods. The work provides a crucial empirical benchmark for selecting appropriate FL strategies for real-world healthcare applications.

arXiv:2509.06924v2 Announce Type: replace Abstract: Neutron reflectometry (NR) is a powerful technique to probe surfaces and interfaces. NR is inherently an indirect measurement technique, access to the physical quantities of interest (layer thickness, scattering length density, roughness), necessitate the solution of an inverse modelling problem, that is inefficient for large amounts of data or complex multiplayer structures (e.g. lithium batteries / electrodes). Recently, surrogate machine learning models have been proposed as an alternative to existing optimisation routines. Although such approaches have been successful, physical intuition is lost when replacing governing equations with fast neural networks. Instead, we propose a novel and efficient approach; to optimise reflectivity data analysis by performing gradient descent on the forward reflection model itself. Herein, automatic differentiation techniques are used to evaluate exact gradients of the error function with respect to the parameters of interest. Access to these quantities enables users of neutron reflectometry to harness a host of powerful modern optimisation and inference techniques that remain thus far unexploited in the context of neutron reflectometry. This paper presents two benchmark case studies; demonstrating state-of-the-art performance on a thick oxide quartz film, and robust co-fitting performance in the high complexity regime of organic LED multilayer devices. Additionally, we provide an open-source library of differentiable reflectometry kernels in the python programming language so that gradient based approaches can readily be applied to other NR datasets.

arXiv:2508.18982v2 Announce Type: replace Abstract: Time series forecasting has seen considerable improvement during the last years, with transformer models and large language models driving advancements of the state of the art. Modern forecasting models are generally opaque and do not provide explanations for their forecasts, while well-known post-hoc explainability methods like LIME are not suitable for the forecasting context. We propose PAX-TS, a model-agnostic post-hoc algorithm to explain time series forecasting models and their forecasts. Our method is based on localized input perturbations and results in multi-granular explanations. Further, it is able to characterize cross-channel correlations for multivariate time series forecasts. We clearly outline the algorithmic procedure behind PAX-TS, demonstrate it on a benchmark with 7 algorithms and 10 diverse datasets, compare it with two other state-of-the-art explanation algorithms, and present the different explanation types of the method. We found that the explanations of high-performing and low-performing algorithms differ on the same datasets, highlighting that the explanations of PAX-TS effectively capture a model's behavior. Based on time step correlation matrices resulting from the benchmark, we identify 6 classes of patterns that repeatedly occur across different datasets and algorithms. We found that the patterns are indicators of performance, with noticeable differences in forecasting error between the classes. Lastly, we outline a multivariate example where PAX-TS demonstrates how the forecasting model takes cross-channel correlations into account. With PAX-TS, time series forecasting models' mechanisms can be illustrated in different levels of detail, and its explanations can be used to answer practical questions on forecasts.

arXiv:2508.16634v2 Announce Type: replace Abstract: Few-Shot Class-Incremental Fault Diagnosis (FSC-FD), which aims to continuously learn from new fault classes with only a few samples without forgetting old ones, is critical for real-world industrial systems. However, this challenging task severely amplifies the issues of catastrophic forgetting of old knowledge and overfitting on scarce new data. To address these challenges, this paper proposes a novel framework built upon Dual-Granularity Representations, termed the Dual-Granularity Guidance Network (DGGN). Our DGGN explicitly decouples feature learning into two parallel streams: 1) a fine-grained representation stream, which utilizes a novel Multi-Order Interaction Aggregation module to capture discriminative, class-specific features from the limited new samples. 2) a coarse-grained representation stream, designed to model and preserve general, class-agnostic knowledge shared across all fault types. These two representations are dynamically fused by a multi-semantic cross-attention mechanism, where the stable coarse-grained knowledge guides the learning of fine-grained features, preventing overfitting and alleviating feature conflicts. To further mitigate catastrophic forgetting, we design a Boundary-Aware Exemplar Prioritization strategy. Moreover, a decoupled Balanced Random Forest classifier is employed to counter the decision boundary bias caused by data imbalance. Extensive experiments on the TEP benchmark and a real-world MFF dataset demonstrate that our proposed DGGN achieves superior diagnostic performance and stability compared to state-of-the-art FSC-FD approaches. Our code is publicly available at https://github.com/MentaY/DGGN

arXiv:2508.16161v2 Announce Type: replace Abstract: Spatio-temporal tasks often encounter incomplete data arising from missing or inaccessible sensors, making spatio-temporal kriging crucial for inferring the completely missing temporal information. However, current models struggle with ensuring the validity and generalizability of inferred spatio-temporal patterns, especially in capturing dynamic spatial dependencies and temporal shifts, and optimizing the generalizability of unknown sensors. To overcome these limitations, we propose Spatio-Temporal Aware Graph Adversarial Neural Network (STA-GANN), a novel GNN-based kriging framework that improves spatio-temporal pattern validity and generalization. STA-GANN integrates (i) Decoupled Phase Module that senses and adjusts for timestamp shifts. (ii) Dynamic Data-Driven Metadata Graph Modeling to update spatial relationships using temporal data and metadata; (iii) An adversarial transfer learning strategy to ensure generalizability. Extensive validation across nine datasets from four fields and theoretical evidence both demonstrate the superior performance of STA-GANN.

arXiv:2508.07208v2 Announce Type: replace Abstract: In-context learning (ICL) is a hallmark capability of transformers, through which trained models learn to adapt to new tasks by leveraging information from the input context. Prior work has shown that ICL emerges in transformers due to the presence of special circuits called induction heads. Given the equivalence between induction heads and conditional k-grams, a recent line of work modeling sequential inputs as Markov processes has revealed the fundamental impact of model depth on its ICL capabilities: while a two-layer transformer can efficiently represent a conditional 1-gram model, its single-layer counterpart cannot solve the task unless it is exponentially large. However, for higher order Markov sources, the best known constructions require at least three layers (each with a single attention head) - leaving open the question: can a two-layer single-head transformer represent any kth-order Markov process? In this paper, we precisely address this and theoretically show that a two-layer transformer with one head per layer can indeed represent any conditional k-gram. Thus, our result provides the tightest known characterization of the interplay between transformer depth and Markov order for ICL. Building on this, we further analyze the learning dynamics of our two-layer construction, focusing on a simplified variant for first-order Markov chains, illustrating how effective in-context representations emerge during training. Together, these results deepen our current understanding of transformer-based ICL and illustrate how even shallow architectures can surprisingly exhibit strong ICL capabilities on structured sequence modeling tasks.

arXiv:2507.18320v2 Announce Type: replace Abstract: The rapid adoption of battery-powered vehicles and energy storage systems over the past decade has made battery health monitoring increasingly critical. Batteries play a central role in the efficiency and safety of these systems, yet they inevitably degrade over time due to repeated charge-discharge cycles. This degradation leads to reduced energy efficiency and potential overheating, posing significant safety concerns. Accurate estimation of a State of Health (SoH) of battery is therefore essential for ensuring operational reliability and safety. Several machine learning architectures, such as LSTMs, transformers, and encoder-based models, have been proposed to estimate SoH from discharge cycle data. However, these models struggle with the irregularities inherent in real-world measurements: discharge readings are often recorded at non-uniform intervals, and the lengths of discharge cycles vary significantly. To address this, most existing approaches extract features from the sequences rather than processing them in full, which introduces information loss and compromises accuracy. To overcome these challenges, we propose a novel architecture: Time-Informed Dynamic Sequence Inverted Transformer (TIDSIT). TIDSIT incorporates continuous time embeddings to effectively represent irregularly sampled data and utilizes padded sequences with temporal attention mechanisms to manage variable-length inputs without discarding sequence information. Experimental results on the NASA battery degradation dataset show that TIDSIT significantly outperforms existing models, achieving over 50% reduction in prediction error and maintaining an SoH prediction error below 0.58%. Furthermore, the architecture is generalizable and holds promise for broader applications in health monitoring tasks involving irregular time-series data.

arXiv:2507.17984v2 Announce Type: replace Abstract: Data-driven traffic state estimation and prediction (TSEP) relies heavily on data sources that contain sensitive information. While the abundance of data has fueled significant breakthroughs, particularly in machine learning-based methods, it also raises concerns regarding privacy, cybersecurity, and data freshness. These issues can erode public trust in intelligent transportation systems. Recently, regulations have introduced the "right to be forgotten", allowing users to request the removal of their private data from models. As machine learning models can remember old data, simply removing it from back-end databases is insufficient in such systems. To address these challenges, this study introduces a novel learning paradigm for TSEP-Machine Unlearning TSEP-which enables a trained TSEP model to selectively forget privacy-sensitive, poisoned, or outdated data. By empowering models to "unlearn," we aim to enhance the trustworthiness and reliability of data-driven traffic TSEP.

arXiv:2507.11411v3 Announce Type: replace Abstract: Multi-task learning (MTL) has shown effectiveness in exploiting shared information across tasks to improve generalization. MTL assumes tasks share similarities that can improve performance. In addition, boosting algorithms have demonstrated exceptional performance across diverse learning problems, primarily due to their ability to focus on hard-to-learn instances and iteratively reduce residual errors. This makes them a promising approach for learning multi-task problems. However, real-world MTL scenarios often involve tasks that are not well-aligned (known as outlier or adversarial tasks), which do not share beneficial similarities with others and can, in fact, deteriorate the performance of the overall model. To overcome this challenge, we propose Robust-Multi-Task Gradient Boosting (R-MTGB), a novel boosting framework that explicitly models and adapts to task heterogeneity during training. R-MTGB structures the learning process into three sequential blocks: (1) learning shared patterns, (2) partitioning tasks into outliers and non-outliers with regularized parameters, and (3) fine-tuning task-specific predictors. This architecture enables R-MTGB to automatically detect and penalize outlier tasks while promoting effective knowledge transfer among related tasks. Our method integrates these mechanisms seamlessly within gradient boosting, allowing robust handling of noisy or adversarial tasks without sacrificing accuracy. Extensive experiments on both synthetic benchmarks and real-world datasets demonstrate that our approach successfully isolates outliers, transfers knowledge, and consistently reduces prediction errors for each task individually, and achieves overall performance gains across all tasks. These results highlight robustness, adaptability, and reliable convergence of R-MTGB in challenging MTL environments.

arXiv:2507.10834v3 Announce Type: replace Abstract: Assortment optimization seeks to select a subset of substitutable products, subject to constraints, to maximize expected revenue. The problem is NP-hard due to its combinatorial and nonlinear nature and arises frequently in industries such as e-commerce, where platforms must solve thousands of such problems each minute. We propose a graph convolutional network (GCN) framework to efficiently solve constrained assortment optimization problems. Our approach constructs a graph representation of the problem, trains a GCN to learn the mapping from problem parameters to optimal assortments, and develops three inference policies based on the GCN's output. Owing to the GCN's ability to generalize across instance sizes, patterns learned from small-scale samples can be transferred to large-scale problems. Numerical experiments show that a GCN trained on instances with 20 products achieves over 85% of the optimal revenue on problems with up to 2,000 products within seconds, outperforming existing heuristics in both accuracy and efficiency. We further extend the framework to settings with an unknown choice model using transaction data and demonstrate similar performance and scalability.

arXiv:2507.03291v2 Announce Type: replace Abstract: Deep learning-based domain adaptation (DA) methods have shown strong performance by learning transferable representations. However, their reliance on mini-batch training limits global distribution modeling, leading to unstable alignment and suboptimal generalization. We propose Global Variational Inference Enhanced Domain Adaptation (GVI-DA), a framework that learns continuous, class-conditional global priors via variational inference to enable structure-aware cross-domain alignment. GVI-DA minimizes domain gaps through latent feature reconstruction, and mitigates posterior collapse using global codebook learning with randomized sampling. It further improves robustness by discarding low-confidence pseudo-labels and generating reliable target-domain samples. Extensive experiments on four benchmarks and thirty-eight DA tasks demonstrate consistent state-of-the-art performance. We also derive the model's evidence lower bound (ELBO) and analyze the effects of prior continuity, codebook size, and pseudo-label noise tolerance. In addition, we compare GVI-DA with diffusion-based generative frameworks in terms of optimization principles and efficiency, highlighting both its theoretical soundness and practical advantages.

arXiv:2507.01695v2 Announce Type: replace Abstract: Deep neural networks (DNNs) have become ubiquitous thanks to their remarkable ability to model complex patterns across various domains such as computer vision, speech recognition, robotics, etc. While large DNN models are often more accurate than simpler, lightweight models, they are also resource- and energy-hungry. Hence, it is imperative to design methods to reduce reliance on such large models without significant degradation in output accuracy. The high computational cost of these models is often necessary only for a reduced set of challenging inputs, while lighter models can handle most simple ones. Thus, carefully combining properties of existing DNN models in a dynamic, input-based way opens opportunities to improve efficiency without impacting accuracy. In this work, we introduce PERTINENCE, a novel online method designed to analyze the complexity of input features and dynamically select the most suitable model from a pre-trained set to process a given input effectively. To achieve this, we employ a genetic algorithm to explore the training space of an ML-based input dispatcher, enabling convergence towards the Pareto front in the solution space that balances overall accuracy and computational efficiency. We showcase our approach on state-of-the-art Convolutional Neural Networks (CNNs) trained on the CIFAR-10 and CIFAR-100, as well as Vision Transformers (ViTs) trained on TinyImageNet dataset. We report results showing PERTINENCE's ability to provide alternative solutions to existing state-of-the-art models in terms of trade-offs between accuracy and number of operations. By opportunistically selecting among models trained for the same task, PERTINENCE achieves better or comparable accuracy with up to 36% fewer operations.

arXiv:2506.13958v2 Announce Type: replace Abstract: Elastic Decision Transformers (EDTs) have proved to be particularly successful in offline reinforcement learning, offering a flexible framework that unifies sequence modeling with decision-making under uncertainty. Recent research has shown that incorporating intrinsic motivation mechanisms into EDTs improves performance across exploration tasks, yet the representational mechanisms underlying these improvements remain unexplored. In this paper, we introduce a systematic post-hoc explainability framework to analyze how intrinsic motivation shapes learned embeddings in EDTs. Through statistical analysis of embedding properties (including covariance structure, vector magnitudes, and orthogonality), we reveal that different intrinsic motivation variants create fundamentally different representational structures. Our analysis demonstrates environment-specific correlation patterns between embedding metrics and performance that explain why intrinsic motivation improves policy learning. These findings show that intrinsic motivation operates beyond simple exploration bonuses, acting as a representational prior that shapes embedding geometry in biologically plausible ways, creating environment-specific organizational structures that facilitate better decision-making.

arXiv:2506.10680v4 Announce Type: replace Abstract: Continual learning poses a fundamental challenge for neural systems, which often suffer from catastrophic forgetting when exposed to sequential tasks. Self-Organizing Maps (SOMs), despite their interpretability and efficiency, are not immune to this issue. In this paper, we introduce Saturation Self-Organizing Maps (SatSOM)-an extension of SOMs designed to improve knowledge retention in continual learning scenarios. SatSOM incorporates a novel saturation mechanism that gradually reduces the learning rate and neighborhood radius of neurons as they accumulate information. This effectively freezes well-trained neurons and redirects learning to underutilized areas of the map.

arXiv:2506.09105v2 Announce Type: replace Abstract: We present MetaTT, a Tensor Train (TT) adapter framework for fine-tuning of pre-trained transformers. MetaTT enables flexible and parameter-efficient model adaptation by using a single shared TT to factorize transformer sub-modules. This factorization indexes key structural dimensions, including layer and matrix type, and can optionally incorporate heads and tasks. This design allows MetaTT's parameter count to scale with the sum, rather than the product, of the modes, resulting in a substantially more compact adapter. Our benchmarks compare MetaTT with LoRA along with recent state-of-the-art matrix and tensor decomposition based fine-tuning methods. We observe that when tested on single-task standard language modeling benchmarks, MetaTT achieves competitive parameter efficiency to accuracy tradeoff. We further demonstrate that MetaTT performs competitively when compared to state-of-the-art methods on multi-task learning. Finally, we leverage the TT-ansatz to design a rank adaptive optimizer inspired by the DMRG method from many-body physics. Our results demonstrate that integrating this approach with AdamW enhances optimization performance for a specified target rank.

arXiv:2506.08312v2 Announce Type: replace Abstract: Private Evolution (PE) is a promising training-free method for differentially private (DP) synthetic data generation. While it achieves strong performance in some domains (e.g., images and text), its behavior in others (e.g., tabular data) is less consistent. To date, the only theoretical analysis of the convergence of PE depends on unrealistic assumptions about both the algorithm's behavior and the structure of the sensitive dataset. In this work, we develop a new theoretical framework to understand PE's practical behavior and identify sufficient conditions for its convergence. For $d$-dimensional sensitive datasets with $n$ data points from a convex and compact domain, we prove that under the right hyperparameter settings and given access to the Gaussian variation API proposed in \cite{PE23}, PE produces an $(\varepsilon, \delta)$-DP synthetic dataset with expected 1-Wasserstein distance $\tilde{O}(d(n\varepsilon)^{-1/d})$ from the original; this establishes worst-case convergence of the algorithm as $n \to \infty$. Our analysis extends to general Banach spaces as well. We also connect PE to the Private Signed Measure Mechanism, a method for DP synthetic data generation that has thus far not seen much practical adoption. We demonstrate the practical relevance of our theoretical findings in experiments.