Explainable AI (XAI) is essential for building trust in Deep Neural Networks (DNNs). SHAP (SHapley Additive exPlanations) is a well‐known XAI technique for attributing feature importance, but it struggles with exponential computational complexity as the number of features increases. Various approximation methods have been suggested, but they compromise SHAP's theoretical principles. We introduce AA‐SHAP, a novel approach that derives superpixel affinity from the explained model's internals to identify and group superpixels. AA‐SHAP constructs a relevance‐consistency affinity between superpixel interdependence, enabling much faster SHAP calculations on a reduced set of meta‐superpixels while outperforming previous methods in explanation faithfulness. Exact Shapley values are computed on the reduced meta‐superpixel game, preserving all axiomatic guarantees within the aggregated feature space. Evaluated across multiple datasets, both convolutional and transformer classification architectures show that AA‐SHAP produces more faithful attributions than competing methods while improving computational speed and maintaining SHAP's theoretical axioms. The source code is available at https://github.com/vhasic/AA‐SHAP .

Introductory programming courses remain challenging for many students, which motivates educators to adopt gamification to enhance engagement and learning. More recent work explores adaptive gamification, where game elements and task flow are tailored to individual learners. A key requirement for such adaptation is the ability to predict student success on upcoming tasks. Using a dataset of task attempts collected from a gamified introductory programming activity, we examine the predictive value of coarse-grained knowledge components, task difficulty, and dynamic student performance features. The results show that behavioral signals are substantially more informative than task properties: a student's prior success history and their position within a lesson sequence are the strongest predictors of future correctness. Although advanced topics such as file handling and structures are associated with increased failure rates, their impact is secondary to students' evolving engagement patterns. These findings highlight the role of momentum and practice effects in gamified programming environments and suggest that adaptive systems should prioritize real-time learner progression when providing instructional support. Dataset and the code for our experiments is available at https://osf.io/cajby.

Conversational commerce agents that personalize assistance based on a user’s transactional state (cart contents, checkout progress, order completion) must model that state correctly, or downstream adaptive behavior will be misaligned with the user’s actual journey. We call mismatches between an agent’s claims and the observable event history journey hallucinations, and study a lightweight verification framework that reconstructs a minimal transactional user model from execution logs and checks agent claims against deterministic invariants. On 90 real sessions across four foundation models, trace-aware prompting reaches 99.5–100% user-state accuracy at 84–99% coverage, while unconstrained prompting produces unsupported state assertions at rates up to 8.5%. In a between-subjects user study (N = 42), verified responses were judged more trustworthy (p =.008, r =.43), better at reflecting journey understanding (p =.039, r =.32), and more often factually correct (p <.001, r =.56). The framework provides a practical reliability layer for transactional user-state modeling, helping personalization and dialog policies operate on verified, not hallucinated, user states.

This paper proposes ontology-guided reasoning for affordance-based explanations of robot navigation. In human environments, it is not sufficient for a robot to detect that its route is blocked. It must also reason about what nearby objects afford, which state changes are possible, and which of these changes would allow it to continue safely. We address this problem by representing nearby entities, their affordances, affordance states, and qualitative spatial relations in a local affordance ontology and by evaluating hypothetical object--affordance state changes as candidate explanation factors. This yields explanations that are not only semantically grounded but also actionable. We instantiate the approach in a lightweight benchmark centered on a robot librarian scenario and evaluate it on procedurally generated navigation cases. The results show that ontology-guided reasoning identifies relevant explanation factors more accurately than a semantic-only baseline and remains robust as semantic clutter increases. Overall, the paper argues that affordance ontologies can serve not merely as semantic descriptions of the environment, but as reasoning foundations for explainability and reliable robot autonomy.

Explainable AI (XAI) is crucial for fostering human trust in deep neural network (DNN) predictions, particularly in tasks like image classification. Multiple surveys exist on XAI methodologies, however, the practical usability and reproducibility of these methods remain largely unexplored. This paper addresses this gap by conducting a systematic survey of recent XAI papers published in leading computer vision and AI conferences and journals. We categorize these works, identify prevalent datasets and evaluation metrics, and analyse the associated code repositories. Our analysis reveals that almost 95% of the surveyed codebases are research prototypes rather than published releases, and a concerning majority of two‐thirds of them exhibit inconsistencies with their corresponding publications. These findings highlight the challenges in benchmarking new XAI methods against existing ones and explain the slow adoption of state‐of‐the‐art research in real‐world applications. This paper aims to underscore the importance of releasing well‐documented, readily usable code alongside XAI research to foster a more robust and reproducible ecosystem, ultimately facilitating the development and deployment of trustworthy AI systems. The results of this study are presented on an interactive website Interactive‐XAI.

Robots increasingly provide explanations to support transparency in Human-Robot Interaction (HRI), yet users differ widely in how much explanation they prefer and when it is appropriate. We present a lightweight simulation framework in which a robot selects among explanation policies ranging from no explanation to norm-based, preference-based, and a Bayesian Adaptive (BA) policy that learns user preferences online while respecting normative expectations. Using synthetic user archetypes, we evaluate how these policies trade off utility, alignment, explanation cost, and regret. Results show that BA consistently achieves low regret across individual users while maintaining strong utility and alignment across diverse user archetypes. These findings motivate the development of preference-aware, uncertainty-driven explanation mechanisms for robust, adaptive robot communication in heterogeneous HRI settings.

This study explores first-year Electrical Engineering and Computer Science students’ use and perception of artificial intelligence (AI) tools in a programming course, and their preferences for future development. We conducted an anonymous exploratory survey, consisting of items with predefined response options and open-ended items. Responses to the former and open-ended items were analyzed using descriptive statistics and inductive thematic analysis, respectively. Additionally, we clustered students using the Affinity Propagation algorithm based on their expressed preferences for possible improvements of AI tools The findings show that AI tools are not universally effective, with seven student clusters identified based on differing needs and expectations. Students expressed a need for AI tools that offer more detailed error explanations and guidance rather than just delivering correct solutions. The most common concern among students is the provision of correct solutions without adequate explanations of the underlying mistakes, leading to a lack of deeper understanding This study takes an exploratory approach by examining students’ perceptions and preferences for the design and capabilities of AI tools in helping them learn programming. Clustering students by preferences reveals distinct approaches that may be needed for different groups of learners. Given the limited research on such desires or on applying clustering to them, our analysis offers valuable insights into distinct viewpoints that can guide the design of future personalized educational AI tools

The integration of deep learning into symbolic music generation presents new opportunities for emulating artist-specific musical styles. In this paper, we propose a multi-branch Long Short-Term Memory (LSTM) network designed to generate monophonic melodies conditioned on note pitch, duration, and playback, with a focus on stylistic imitation of The Beatles. Unlike existing approaches that model music solely as sequences of pitches, our model processes three distinct streams of musical attributes and learns joint temporal dependencies through a custom architecture. We introduce a structured data representation derived from 193 MIDI files of Beatles songs using the music21 toolkit, extracting pitch and duration features and quantizing them into a format suitable for sequential prediction. Experimental results demonstrate that the model captures artist-specific musical patterns with moderate accuracy across output branches, and a listening test involving 71 participants validates the perceptual plausibility of the generated compositions. Our findings suggest that feature-aware sequence modeling is effective for stylistically informed symbolic music generation, and we discuss limitations and future extensions toward polyphonic modeling and conditional generation.