A revised arXiv paper proposes a graph neural network framework for diagnosing which specific sensor or component is responsible for anomalies in multivariate industrial time series. The work aims to go beyond simple anomaly detection by making the source of the fault interpretable. It targets reliability and safety monitoring in complex, multi-sensor process environments.
A new arXiv preprint describes an approach that derives logical explanations for neural network classifications by analyzing neuron activations. The work situates itself within formal explainability, which aims to give provable guarantees about model behavior across regions of the input space. The abstract notes that existing formal techniques have limitations the proposed method seeks to address, though details of the approach are not included in the announcement.
A new arXiv preprint proposes an explainable screening approach for Parkinson's disease based on hand-drawn spirals and meanders. The work argues that tremor-driven oscillations, irregular strokes, and unstable curvature in these drawings reveal early neuromotor impairment. It then connects visual attribution methods to clinical reasoning so the model's outputs can be interpreted by clinicians.
A new arXiv preprint proposes treating safety as a constraint when fine-tuning a large language model used as a recommender system. Standard recommenders are trained only to predict the next item a user will engage with, not to justify the prediction, so the authors add self-explanation as a training objective. The goal is to give users personalized reasons for suggestions without letting the generated explanations violate safety requirements.
A new arXiv paper presents an approach that combines data storytelling with interpretable machine learning to make model decisions understandable to people without technical backgrounds. The method is designed to explain predictions while avoiding disclosure of sensitive training data or proprietary model internals. The authors position the work as addressing the tension between predictive performance and interpretability in automated decision-making.
A new arXiv paper examines how explainable AI methods are assessed, grouping evaluation approaches into subjective measures such as user trust questionnaires, objective measures based on task performance, and mathematical metrics. The authors analyze whether these three families of evaluation strategies produce consistent conclusions by testing them with saliency maps. The work aims to clarify where the different measurement types agree or diverge.
A new arXiv paper describes an intrusion detection system designed for Internet of Medical Things networks, where devices are diverse and computing power is limited. The approach combines hybrid feature selection with explainability so that real-time traffic can be screened while keeping the model's decisions interpretable. The authors frame resource constraints and the need for timely analysis as the main obstacles the method targets.
A new arXiv paper introduces SAILS, a surrogate-based approach that uses local effect smooths to characterize how features interact inside machine learning models. Unlike prior explanation techniques that only flag or score interactions, or that handle just a narrow set of interaction shapes, SAILS aims to expose the actual functional form of those interactions. The work is posted as a cross-listing on arXiv's cs.AI and cs.LG categories.
An arXiv preprint argues that as 6G moves from theory toward deployment, AI shifts from a bolt-on optimization aid to an interconnected layer woven through the network itself. The authors outline requirements for making that distributed intelligence trustworthy, explainable and sustainable, and sketch an architecture to meet them. The work is a cross-listed submission focused on research directions rather than a working system or product.
A new arXiv preprint describes an approach for generating counterfactual examples that are both varied and human-interpretable, drawing on expert knowledge to guide the search. Counterfactual examples are a common technique in explainable AI, since they show the smallest input changes that would flip a model's prediction. The work aims to address the limited diversity typical of existing methods while keeping the resulting explanations understandable.
A new arXiv paper examines whether deep learning and explainable AI methods can help diagnose hepatocellular carcinoma and pinpoint biomarkers across five stages of disease progression. The work uses a transcriptomic biomarker dataset for liver cancer, aiming for models whose predictions can be traced to interpretable biological features. The authors frame it as an early exploration of combining accuracy with explainability in cancer diagnostics.
A new arXiv paper introduces AI Soccer Analyst, a system designed to help sports analysts work with soccer data through a stage-aware, verifiable human-AI collaboration workflow. The authors argue that while large language models make programming easier for analysts, prompt-to-report pipelines can hide the reasoning steps and supporting evidence behind conclusions. Their approach aims to keep the analysis process transparent and checkable by structuring collaboration around distinct stages.
A new arXiv preprint looks at how neural networks in speaker recognition pick up patterns beyond those explicitly defined by researchers, treating these as "second-order" patterns. The work connects classical pattern recognition training with explainable AI methods that surface latent features underlying a network's decisions. It focuses on the speaker recognition domain as a case study.
Researchers propose a chat-based explainability assistant designed to help building operators and facility managers understand predictions from complex energy consumption models, including symbolic regressors built with genetic programming. The tool is presented as a way to make model outputs more accessible to non-experts who manage energy use.
A new arXiv preprint introduces X-RACE, a method that applies explainable AI attribution techniques to recurrent neural networks, specifically LSTMs, used for channel estimation in high-mobility vehicular settings. The approach aims to address the limited interpretability and computational overhead that hinder trust in deep learning models for this task.
A new arXiv paper argues that existing counterfactual explanation methods for temporal graphs only show how to alter past events to flip a prediction, without stating what the resulting outcome should become. The authors introduce "specified-foil" counterfactuals, which target a particular alternative outcome rather than leaving the replacement undefined. The work aims to give users more actionable explanations of predicted events in time-evolving graph data.
A systematization of knowledge paper reviews 25 studies showing how explainable AI outputs can be turned against the models they describe. The authors group these attacks into model extraction, membership inference and model inversion, and note that explanations widen the confidentiality and privacy risks of deployed systems. The work calls for treating explanation interfaces as part of the attack surface.
A new arXiv paper examines automated simulatability, a protocol that scores explanations by how well they let a user predict a model's outputs without relying on costly human evaluation. The authors report that when LLMs stand in for human explainees, they can bypass the explanations themselves, undermining the validity of the metric. The work suggests automated simulatability may overstate how useful an explanation really is.
A new preprint introduces a decomposition method that extends feature-attribution explanations from single-number predictions to functional or multivariate outputs, such as demand forecasts that vary over time. The approach works in a Hilbert space so that the influence of each input feature can be separated across the whole output trajectory rather than summarized by one score. The authors position it as a general framework for settings where model predictions are curves or vectors instead of scalars.
A new arXiv paper introduces XAI-Arena, a study of whether large language models can reliably evaluate explanations produced by explainable AI methods. The authors note that current evaluation relies heavily on subjective human judgment, which hurts reproducibility, scalability, and comparability across studies. The work explores automated, LLM-based assessment as a potential alternative to manual expert reviews.
A new paper introduces CARRE, a three-stage framework that moves churn analysis beyond simply flagging customers likely to leave. It retrieves feasible counterfactual interventions and evaluates the reasoning behind each recommendation, so that retention suggestions come with a justified rationale rather than a risk score alone.
A new arXiv paper presents an explainable machine learning framework that predicts blood-brain barrier permeability using molecular descriptors. Since this barrier determines whether central nervous system drug candidates can reach targets in the brain, the approach could support earlier screening in drug development. Its explainable design is intended to reveal which molecular features drive the model's predictions.
Researchers introduce LM-X, a framework for vision-language-action robot policies that exposes an explanatory state alongside its actions. Instead of acting as a stimulus-to-action black box, the model natively predicts task progress, notable events, and uncertainty in its decisions. The work aims to bring interpretability to large-scale generalist robot control.
A new arXiv preprint introduces XAI-Refine, a framework that couples post-hoc explanation methods with a knowledge-refinement loop for brain-age prediction models. The authors argue that high predictive accuracy alone does not show a model relies on reproducible or neurobiologically plausible mechanisms, so explanations are fed back to iteratively improve the model. The approach aims to make brain-age estimates more interpretable and mechanistically grounded.
A position paper examines how large language models generate natural-language accounts of their own decisions, a practice known as self-explanation. The authors argue that current explanations are often merely plausible-sounding rather than genuinely useful, and they outline what would be needed to make them actionable for real-world use.