Explainable AI: Unpacking the Black Box Across Domains, from Medical Diagnostics to Cybersecurity

Latest 50 papers on explainable ai: Dec. 13, 2025

The quest for intelligent systems has never been more fervent, but as AI models grow in complexity, the demand for transparency and trustworthiness has surged. Enter Explainable AI (XAI) – a rapidly evolving field dedicated to making these powerful black boxes understandable to humans. Recent research has pushed the boundaries of XAI, delivering groundbreaking innovations that empower users across diverse domains, from healthcare to cybersecurity and industrial automation. This digest delves into the latest advancements, highlighting how XAI is transforming theoretical concepts into actionable, real-world solutions.

The Big Idea(s) & Core Innovations

One of the most compelling trends is the drive towards actionable and human-centric explanations. Researchers from the University of California, San Francisco and Stanford University introduce Motion2Meaning: A Clinician-Centered Framework for Contestable LLM in Parkinson’s Disease Gait Interpretation, a framework that allows clinicians to challenge and refine AI predictions, boosting trust in medical AI. Complementing this, J. Shymanski et al. from The University of Tulsa (Cyber Fellows program), in Beyond Satisfaction: From Placebic to Actionable Explanations For Enhanced Understandability, argue that user satisfaction alone is insufficient for XAI, advocating for objective performance measures for actionable explanations that truly enhance understanding.

Driving interpretability deeper, Jules Soria et al. from Université Paris-Saclay, CEA, List propose Formal Abductive Latent Explanations for Prototype-Based Networks. Their Abductive Latent Explanations (ALEs) offer rigorous, scalable, and trustworthy explanations at the latent level, moving beyond pixel-based interpretations for prototype-based networks. Similarly, Bui Tien Cuong from Seoul National University, in Enhancing Explainability of Graph Neural Networks Through Conceptual and Structural Analyses and Their Extensions, introduces a novel framework for Graph Neural Networks (GNNs) that combines conceptual and structural analyses to provide computationally efficient and user-centric explanations.

Beyond just understanding, XAI is now actively improving model performance. The π3Lab, Peking University team, in IVY-FAKE: A Unified Explainable Framework and Benchmark for Image and Video AIGC Detection, presents Ivy-xDetector, which leverages reinforcement learning to produce fine-grained explanations for AI-generated content (AIGC) detection, achieving state-of-the-art performance. Furthermore, Seoyeon Lee et al. from Kookmin University, in Refining Visual Artifacts in Diffusion Models via Explainable AI-based Flaw Activation Maps, show how XAI-based Flaw Activation Maps (FAMs) can actively refine image generation in diffusion models, demonstrating that XAI isn’t just for post-hoc analysis but can be integral to the generation process itself. This self-refining approach significantly improves output quality across diverse tasks.

Another critical area is making AI robust and compliant. Dr. Swati Sachan from University of Liverpool and Prof. Dale S. Fickett from Robins School of Business, University of Richmond, in DeFi TrustBoost: Blockchain and AI for Trustworthy Decentralized Financial Decisions, integrate blockchain with XAI to provide tamper-proof auditing for automated financial decisions, enhancing trust and regulatory compliance. Moreover, Anton Hummel et al. (XITASO GmbH & University of Bayreuth) bridge the gap between XAI and the EU AI Act for clinical decision support systems, emphasizing that XAI is crucial for regulatory alignment and human oversight.

Under the Hood: Models, Datasets, & Benchmarks

The advancements detailed across these papers rely on a combination of innovative models, tailored datasets, and robust benchmarks:

• CIP-Net: Federico Di Valerio et al. from Sapienza University (CIP-Net: Continual Interpretable Prototype-based Network) introduces an exemplar-free, self-explainable continual learning model that uses shared prototype layers for knowledge sharing and targeted regularization to prevent catastrophic forgetting. Code available at https://github.com/KRLGroup/CIP-Net.
• QCAI & TCR-XAI: Jiarui Li et al. from Tulane University (Quantifying Cross-Attention Interaction in Transformers for Interpreting TCR-pMHC Binding) developed QCAI for interpreting cross-attention in transformer decoders and introduced TCR-XAI, a benchmark of experimentally determined TCR-pMHC structures for quantitative XAI evaluation.
• IT-SHAP: Hiroki Hasegawa and Yukihiko Okada from University of Tsukuba (Interaction Tensor Shap) present IT-SHAP, a novel method for efficiently computing high-order feature interactions using tensor-network contractions under TT structure, transforming exponential computational complexity into polynomial time.
• EXCAP: Ziqian Wang et al. from Tsinghua University (A Self-explainable Model of Long Time Series by Extracting Informative Structured Causal Patterns) propose EXCAP, a unified neural framework combining attention-based segmentation, causal decoding, and latent aggregation for self-explainable long time series modeling.
• CONFETTI: For multivariate time series classification, Alan G. Paredes Cetina et al. from SnT, University of Luxembourg (Counterfactual Explainable AI (XAI) Method for Deep Learning-Based Multivariate Time Series Classification) introduce CONFETTI, a multi-objective counterfactual explanation method that balances prediction confidence, proximity, and sparsity. Code available at https://github.com/serval-uni-lu/confetti.
• FunnyNodules: Luisa Gallée et al. from Ulm University Medical Center (FunnyNodules: A Customizable Medical Dataset Tailored for Evaluating Explainable AI) developed FunnyNodules, a fully parameterized synthetic medical dataset to systematically evaluate XAI models by capturing both diagnostic labels and their underlying reasoning. Code available at https://github.com/XRad-Ulm/FunnyNod3456789012345678901234567890123456789012345678901234567890123456789012345678901234567890123456789012.
• XAI-on-RAN: Osman Tugay Basaran and Falko Dressler from Technische Universität Berlin (XAI-on-RAN: Explainable, AI-native, and GPU-Accelerated RAN Towards 6G) propose XAI-on-RAN, a 6G RAN architecture integrating AI-native control with real-time XAI, leveraging GPU acceleration for fast, interpretable decisions.

Impact & The Road Ahead

These advancements signify a profound shift in how we build and interact with AI. XAI is moving beyond mere post-hoc analysis to become an intrinsic part of the AI lifecycle, from design and training to deployment and continuous refinement. The ability to generate actionable explanations will foster greater trust and adoption in high-stakes fields like healthcare, where systems like DeepGI for gastrointestinal image classification (DeepGI: Explainable Deep Learning for Gastrointestinal Image Classification) and the stroke risk prediction pipeline with ROS-balanced ensembles (Optimizing Stroke Risk Prediction: A Machine Learning Pipeline Combining ROS-Balanced Ensembles and XAI) provide not just diagnoses but also transparent reasoning. In cybersecurity, interpretable ransomware detection using LLMs (Interpretable Ransomware Detection Using Hybrid Large Language Models) and XAI-driven backdoor unlearning monitoring (Illuminating the Black Box: Real-Time Monitoring of Backdoor Unlearning in CNNs via Explainable AI) are crucial for robust and auditable defense mechanisms.

Looking ahead, the integration of XAI into emerging technologies like 6G networks with XAI-on-RAN demonstrates its potential to enable trustworthiness in mission-critical communications. The push for certified monotonicity in models like MonoKAN (MonoKAN: Certified Monotonic Kolmogorov-Arnold Network) signifies a growing demand for AI systems that are not just accurate but also predictably fair and transparent. The ongoing research into nonlinear explainability with methods like SISR (Beyond Additivity: Sparse Isotonic Shapley Regression toward Nonlinear Explainability) promises even more nuanced and accurate attributions for complex models. As AI continues to permeate every aspect of our lives, XAI will be the cornerstone of responsible, ethical, and truly intelligent systems.