Reliable and Trustworthy Artificial Intelligence 2025 Held in Multiple on 12 December 2025 Published as Volume 310 by the Proceedings of Machine Learning Research on 31 December 2025. Volume Edited by: Hoang D. Nguyen Duc-Trong Le Johanna Björklund Xuan-Son Vu Series Editors: Neil D. Lawrence https://proceedings.mlr.press/v310/ Wed, 07 Oct 2026 11:43:51 +0000 Wed, 07 Oct 2026 11:43:51 +0000 Jekyll v3.10.0 Integral privacy, an alternative to k-Anonymity and differential privacy, focuses on creating ambiguity for intruders by considering models generated from diverse datasets as privacy-preserving. Integral privacy calls such models as recurring models. While prior research has primarily explored recurrence in deep learning models which have large parameter space, this paper addresses the recurrence analysis of a typical machine learning model with relatively small parameter space like Support Vector Machine (SVM). Models having small parameter space can have significant impact due to the presence and absence of a datapoint. Due to this reason, one may intuitively consider that their probability to recur is low. We challenge this hypothesis with the recurrence analysis of SVM models trained with mean samplers like stochastic gradient descent. We find that under constrained environment SVM models recurs with high probability. This research enhances our understanding of privacy-preserving models in the context of SVMs, providing valuable insights into their privacy guarantees. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/varshney25a.html https://proceedings.mlr.press/v310/varshney25a.html Safety is an integral part of healthcare professionalism, and with new technological developments, such as Artificial Intelligence (AI), there is an ongoing need to develop guardrails for healthcare education. The landscape of AI safety frameworks for healthcare education is evolving, with significant development in regulatory compliance, ethical governance, and practical implementation approaches. This paper addresses the need for building a SAFETY-AI framework for healthcare education and proposes a solution towards it. It also provides subjective insights regarding trustworthiness, reliability and the existing concepts of safety in healthcare setups. This work stands as a roadmap for safety in AI practices for healthcare policy makers, educators and clinicians. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/salim25a.html https://proceedings.mlr.press/v310/salim25a.html Continual learning (CL) refers to the capability of a model to learn progressively from an evolving stream of data, retaining previously acquired knowledge while integrating new information. This capability is pivotal in advancing medical image classification, especially when data availability fluctuates. Beyond investigating CL performance under standard clean-data conditions, this paper systematically evaluates the robustness of representative CL strategies in uncertain imaging contexts, where visual quality is degraded by varying degrees of low-light conditions and over- or under exposure. In this paper, we augment the training and evaluation data with controlled, simulated low-light and contrast perturbations to model these uncertain conditions, which mimic real-world variability frequently encountered in clinical acquisition environments. Our method integrates an automatic illumination calibration module, termed the Illumination Adaptive Transformer (IAT), within existing CL frameworks to mitigate the adverse effects of such degradations. This module dynamically adjusts the image illumination and contrast, aiming to enhance the visibility of critical features in a data-driven, end-to-end manner without requiring manual tuning or image-specific heuristics. Experiments demonstrate that incorporating the IAT module consistently improves final classification accuracy and robustness across multiple continual learning strategies under all simulated uncertainty levels on the PathMNIST dataset. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/phan25a.html https://proceedings.mlr.press/v310/phan25a.html The nexus between data characteristics and parametric models is fundamental for developing effective and reliable artificial intelligence (AI) systems. Mismatches in data properties for model development may lead to deleterious effects on AI model performance in machine learning practice. This paper proposes a Reliable Data Split (RDS) procedure to learn how to select data points that will generalise the target domain adequately by employing prior knowledge of the data generative process. We introduce a reinforced selection strategy using deep reinforcement learning with diverse black box predictors in maximising ensemble rewards as the proxy of model performance potential while maintaining an appropriate proportionate allocation and the independent and identically distributed (i.i.d.) assumption. A comprehensive evaluation of the RDS procedure is conducted on four real-world datasets, including Madelon, Drug Reviews, MNIST, and Kalapa Credit Scoring Challenge, with coverage of machine learning tasks such as binary classification, multi-class classification, and regression on multivariate, textual, and visual data. The experimental results evidently demonstrate consistent performance improvements of trainable data samples over classical or prior data selection. Hence, we advocate the use of RDS for data splitting in the early stage of machine learning tasks for parameter tuning, model selection and overfitting prevention, as well as, sampling in large-scale AI competitions for searching the best possible and shift-stable solutions. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/nguyen25c.html https://proceedings.mlr.press/v310/nguyen25c.html We introduce a reliable approach for enhancing multilingual language models that preserves cultural knowledge while improving reasoning capabilities, focusing on low-resource languages. Using Qwen as a base model, we demonstrate that trust-aware model merging can verifiably improve performance without compromising cultural understanding. Our proposed approach achieves quantifiable improvements in both reasoning tasks and cultural benchmarks while maintaining computational efficiency. Results on Vietnamese and Arabic language tasks show consistent performance gains while preserving cultural knowledge, offering a reliable path for developing trustworthy multilingual AI systems. Our models are available at github.com/WARA-ML/waraml-mini-brains. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/nguyen25b.html https://proceedings.mlr.press/v310/nguyen25b.html Preface Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/nguyen25a.html https://proceedings.mlr.press/v310/nguyen25a.html It is important to do dimensionality reduction or feature selection so that machine learning models can be built in an efficient and interpretable way, especially with high-dimensional datasets. Using data up to October 2023, this study compares two methods, namely, Ge- netic Algorithm (GA) and Principal Component Analysis (PCA), to assess their usefulness for dimensionality reduction while preserving predictive performance. As a case study, we applied the feature selection methodology on the Breast Cancer Wisconsin (Diagnostic) dataset, comprised of 30 real-valued features that describe the characteristics of cell nuclei. PCA used to reduce dimensional space of dataset by explaining 95% of variance and GA is used to make a minimal subset of subset of relevant features based on fitness function. To evaluate the effect of the reduced dimensionality on classification accuracy, a Random Forest classifier was used. Experimental results shown that GA selected features provide accuracy from the GA by 98.25% and PCA accuracy from the PCA with 93.86% which at the cost of high computational cost. Finally, visualizations illustrated the variance re- tained by PCA, how features provided importance in model performance using GA, and how both influence the models. The quantitative results of this study can be used to iden- tify the trade-off between the statistical approaches and heuristic approaches showing what needs to be prioritized in terms of application specificity when searching for dimensionality reduction methods. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/naeem25a.html https://proceedings.mlr.press/v310/naeem25a.html The integration of artificial intelligence (AI) into multi-unmanned aerial vehicle-assisted communication plays a pivotal role in sixth-generation wireless communication and be- yond. Most AI techniques have primarily focused on AI-based applications and technical problems, rather than examining the accountability and trustworthiness of AI models, a crucial evaluation criteria for AI human beings. This work aims to provide a scoping re- view of the trustworthiness of AI in multi-agent UAV systems. Firstly, we present the background of multi-agent systems and methods to evaluate, enhance the trustworthiness of AI systems. Secondly, we review innovative techniques that address trustworthy require- ments in terms of safety, robustness, privacy, accountability, explainability, and fairness, along with challenges in multi-agent UAV communications. Finally, we highlight several promising solutions and future research directions. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/le25a.html https://proceedings.mlr.press/v310/le25a.html Learning analytics can enable the provision of meaningful feedback based on the collected data, help educators to make decisions with and about learners, and improve learner performance. Student engagement predictions are a key factor in generating feedback for real-time learning analytics applications, such as dashboards. However, most previous work has been based on a heavy deep learning model, which results in challenges for deployment in real-time applications (a resource efficiency requirement in reliable AI). This paper proposes a lightweight deep-learning framework for predicting student engagement in video to address this limitation. The proposed method uses customized MobileNetV2 as the backbone, with an input size of 32 by 32 by 3, to extract features from consecutive video frames. Multi-Scale attention – Residual (MUSER) is used to capture global information and contextual representation of the extracted features. Finally, LSTM examines the temporal variations in video frames and yields the prediction result. We use the DAISEE dataset, the most popular dataset in the learning analytics community, to evaluate the proposed framework. Experimental results demonstrate that the proposed method achieves good accuracy while significantly reducing the model size compared to other approaches. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/hoang25a.html https://proceedings.mlr.press/v310/hoang25a.html The rapid deployment of large language models (LLMs) in critical domains demands greater transparency and explainability to build user trust and enable effective collaboration. Cur- rent AI-human interactions largely rely on unstructured text, often resulting in misunder- standings and limited insight into AI reasoning. We introduce a Hierarchical Conceptual Graph Model and the Knowing Communication Protocol to bridge the gap between sym- bolic human reasoning and sub-symbolic AI processing. Our model combines conceptual spaces, ontologies, and hierarchical structures to explicitly represent complex knowledge, while the Knowing Protocol, through the Knowing Markup Language (KML), facilitates structured, machine-readable interactions. This approach enhances transparency by align- ing AI-generated content with human cognitive structures, promoting clarity and collabo- rative knowledge building—ultimately addressing the limitations of traditional text-based AI tools and advancing trustworthy, explainable AI. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/ehnemark25a.html https://proceedings.mlr.press/v310/ehnemark25a.html The reliability of AI models typically depends on the data they are trained with, and accurate interpretations require large amounts of data. The scarcity of publicly available datasets is typically encountered for specific small-scale sustainability projects, making data accessibility a limiting factor for developing AI models for semantic segmentation tasks. In sustainability and forestry applications, the usage of UAVs is common due to their lightweight nature and the ability to provide a huge variety of data. In this paper, we present a new dataset of realistic and high-quality drone images taken around sites in Ireland. The images encompass temporal, spatial, and seasonal dimensions, which could alter the tree appearance or illumination conditions of the images and have to be taken into consideration. We also included a baseline benchmark for the semantic segmentation task along with the dataset. It can be accessed at: https://github.com/ReML-AI/AIQTrees. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/chai25a.html https://proceedings.mlr.press/v310/chai25a.html The central decision-making processes involved in online advertising are often supported by Reinforcement Learning (RL), which serves to optimise long-term accumulative re- wards through interactions with evolving environments. While RL’s potential in various real-world applications has been reviewed in extant survey works, the specific ways RL algorithms address online advertising challenges remain unchartered. Therefore, this paper reviews RL applications in this practice area, identifying core challenges and key issues including trust concerns. We categorize reviewed work based on problem domains and propose potential directions for future research. Our goal is to bridge the cross-disciplinary gap in this field, offering perspectives and guidance for researchers and practitioners. Wed, 31 Dec 2025 00:00:00 +0000 https://proceedings.mlr.press/v310/cai25a.html https://proceedings.mlr.press/v310/cai25a.html