Research Papers

Explainable AI (XAI) has achieved remarkable success in image classification, yet the audio domain lacks equally mature solutions. Current methods apply vision-based attribution techniques to spectrograms, overlooking fundamental differences between visual and acoustic signals. While prototype reasoning is promising, acoustic similarity remains multidimensional. We introduce APEX (Audio Prototype EXplanations), a post-hoc framework for interpreting pre-trained audio classifiers. Crucially, APEX requires no fine-tuning of the original backbone and strictly preserves output invariance. APEX disentangles explanations into four perspectives: Square-based prototypes to localize transient events, Time-based for temporal patterns, Frequency-based highlighting spectral bands, and Time-Frequency-based integrating both. This yields intuitive, example-based explanations that respect acoustic properties, providing greater semantic clarity than standard gradient-based methods.

Neural networks have proved an effective means of learning control policies for autonomous systems, but these learned policies are difficult to understand due to the black-box nature of neural networks. This lack of interpretability makes safety assurance for such autonomous systems challenging. The fields of eXplainable Artificial Intelligence (XAI) and eXplainable Reinforcement Learning (XRL) aim to interpret the decision making processes of neural networks and autonomous agents, respectively. In particular, work on causal explanations aims to provide "why" and "why not" explanations for why a model made a given decision. However, most of the work on explainability to date utilises a distilled version of the original model. While this distilled policy is interpretable, it necessarily degrades in performance significantly when compared to the original model, and is not guaranteed to be an accurate reflection of the decision making processes in the original model and as such cannot be used to guarantee its safety. Recent work on understanding the geometry of ReLU neural networks shows that a ReLU network corresponds to a piecewise linear function divided into regions defined by an n-dimensional convex polytope. Through this lens, a neural network can be understood as dividing the input space into distinct regions which apply a single linear function for each output neuron. We show that this geometric representation can be used to generate causal explanations for the network's behaviour similar to previous work, but which extracts rules directly from the geometry of Neural Networks with the ReLU activation function, and is therefore an accurate reflection of the network's behaviour.

Open Radio Access Networks (O-RAN) are increasingly adopting data-driven control through Deep Reinforcement Learning (DRL) to optimize complex tasks such as network slicing and mobility management. However, the deployment of DRL in carrier-grade networks is hindered by its inherent opacity and stochastic execution, which limit operator trust, auditability, and safe deployment. Existing explainable AI (XAI) approaches primarily provide post-hoc insights and fail to produce executable, interpretable policies suitable for operational environments. In this paper, we present DeRAN, a neuro-symbolic framework that bridges the gap between DRL performance and operational transparency by distilling black-box DRL policies into human-readable symbolic representations. DeRAN introduces a concept-driven abstraction layer that transforms high-dimensional network telemetry into a compact set of semantically meaningful features, enabling interpretable policy learning. Building on the semantically grounded concepts, DeRAN synthesizes symbolic policies using deep symbolic regression (DSR) for continuous control and neurally guided differentiable logic (NUDGE) for discrete decision-making. We implement DeRAN on a live 5G O-RAN testbed and evaluate it on two representative use cases. Experimental results demonstrate that DeRAN achieves 78% and 87% of DRL's cumulative rewards in the two use cases, while offering interpretability and auditability by design. Source code is available at https://github.com/Jadejavu/DeRAN

As LLMs rapidly advance and become increasingly agentic, mitigating social bias and ensuring fairness at scale have become critical challenges. Prior work has shown that LLMs exhibit an emergent self-debiasing capability-an intrinsic mechanism that enables them to avoid generating biased outputs-but this mechanism remains poorly understood, leaving existing bias-mitigation approaches heavily reliant on carefully curated prompts or inefficient training procedures that risk catastrophic forgetting. To address this gap, we propose I-CAN-Debias, a unified framework for uncovering and enhancing self-debiasing mechanisms in LLMs through behavior-grounded Consistency-contrast-Aware Neuron analysis, which defines self-debiasing neurons not by the magnitude of their responses to biased inputs but by activation patterns that distinguish successful debiasing from biased failures. Using I-CAN-Debias, we find that self-debiasing neurons account for less than 2% of model parameters yet are causally critical: deactivating them raises the biased-response rate above 95%. These neurons cluster in the query and value projections of later layers and exhibit highly asymmetric cross-category transferability. Furthermore, targeted fine-tuning of these neurons substantially improves fairness while preserving general capabilities. Importantly, models tuned with I-CAN-Debias generalize robustly across diverse evaluations beyond the training distribution, spanning open-ended fairness, jailbreak attacks, and implicit bias, without requiring test-time debiasing prompts. Together, these findings provide new mechanistic insights and a parameter-efficient approach toward developing fair and trustworthy AI systems.

We establish, from the point of view of Explainable AI (XAI), connections between Consistency-Based Diagnosis (CBD), on one side, and Actual Causality and Causal Responsibility, on the other. CBD has received little attention from the XAI community. Connections between these two areas could have a fruitful impact on XAI and Explainable Data Management.

Ante-hoc interpretability methods based on prototypes provide highly accurate explanations by utilizing the intuitive "this looks like that" reasoning paradigm. On the other hand, post-hoc models can explain predictions for a single image without relying on an underlying dataset or requiring costly neural network retraining. Recent approaches successfully solve the retraining problem for prototype-based networks. However, they still face a fundamental limitation: they require access to a subset of data (e.g., a test or validation set) to search for and extract the visual prototypes. In this paper, we address this issue and introduce ProDG: Generative Prototypes for Data-Free Post-Hoc Explainability, a novel framework that leverages generative models to synthesize pure, high-fidelity prototypes directly from the frozen model's weights, completely eliminating the dependency on any external data. By establishing this new frontier in Data-Free XAI, ProDG unlocks robust visual interpretability for privacy-sensitive domains, where original data is strictly restricted or fundamentally inaccessible. Project page: https://github.com/piotr310100/ProDG

The current catalyst discovery and development pipeline for energy-intensive applications like methane conversion remains bottlenecked by expensive trial-and-error experimentation, irreproducible chemical intuition, and a lack of frameworks linking complex catalytic design spaces to performance. This work presents an interpretable machine learning framework that integrates SHAP-based feature importance analysis (Explainable AI) with tree-based ensembles (Random Forest and Bayesian-optimized CatBoost) to characterize Fe-zeolite and oxide-supported catalysts for the partial oxidation of methane (POM). Despite limited data, the framework decodes complex structure-performance relationships by identifying and ranking thermodynamic, structural, and geometric microdescriptors that influence the electronic band gap and govern macroscale performance metrics such as selectivity, activity, and stability. This work explicitly demonstrates that thermodynamic lattice stability and geometric factors are the primary drivers of electronic band gap (a critical proxy for redox reactivity) rather than bulk stoichiometry. Non-linear models achieve an R2 of 0.61 - 0.77, significantly outperforming traditional linear baselines (R2 = 0.32). This workflow provides both a light-weight generalizable methodology and a prioritized list of physical features for accelerated catalyst screening - and these features can subsequently be integrated into microkinetic and reaction engineering models to create digital twins of complex reactor systems and to enable predictive optimization in autonomous R&D laboratories.

Machine learning-based Android malware detectors often fail in real-world deployment due to domain shift, where models trained on one data source perform poorly on applications from another. This paper presents a comprehensive study on the generalizability and interpretability of permission-based detectors under cross-domain conditions. Using two complementary datasets (PerMalDroid and NATICUSdroid) and five ensemble classifiers, we first establish an intra-domain baseline, where models achieve over 92% accuracy, and then quantify a severe asymmetric performance drop. While models trained on PerMalDroid generalize well to NATICUSdroid (86% accuracy), the reverse direction sees a drastic drop to 73% accuracy. Explainable AI analysis reveals bimodal feature distributions and shows that feature importance is highly unstable, with key permissions losing or gaining influence across domains. The predictive feature sets for different domains are fundamentally mismatched, as models rely on different, dataset-specific permissions. Most importantly, an ablation study demonstrates that for most models, training on a noisy feature set leads to poor generalization, confirming that domain-specific artifacts are a greater obstacle than missing features. To mitigate this, we validate a hybrid training strategy based on the intersection of common features and successfully recover cross-domain performance, achieving 88% accuracy on PerMalDroid and maintaining 97% on NATICUSdroid. These findings highlight the importance of explainable, cross-domain-robust malware detection systems and provide a practical pathway toward improving real-world deployment of permission-based Android malware detectors.

Reliable short horizon forecasting of solar and wind generation is a structural prerequisite of any modern power system yet most published forecasters are tuned and evaluated on a single climatic regime and most algorithmic novelty has been concentrated either on classical recurrent networks or on monolithic foundation models that combine forecasting and explanation We develop a four stage hybrid framework that separates these concerns The first stage acquires hourly generation irradiance and surface weather records through public application programming interfaces The second stage trains three classical baselines autoregressive integrated moving average gradient boosted regression trees and a two layer long short term memory network and produces a strong point forecast together with a residual error series The third stage corrects the residual through a quantum inspired variational kernel built on a six qubit hardware efficient ansatz with three repeated entangling layers The fourth stage uses generative artificial intelligence strictly as an explainability layer that reads the measured benchmark numbers and produces a structured natural language interpretation Across three regions drawn from open public archives Iberian solar North Sea wind and a mixed Texas trace the proposed configuration stays within one percentage point of the strongest classical baseline on the in domain forecasting task and the quantum inspired kernel separates calm and stormy weather regimes with a Fisher discriminant ratio approximately fifteen fold higher than a tuned radial basis kernel

Investigating Anisotropy in Visual Grounding under Controlled Counterfactual Perturbations

Gabriele Lombardo, Luigi Maiorana, Liliana Lo Presti +1

Visual Grounding benchmarks assume that the object described by a referring expression is always present in the image, and grounding models are therefore rarely evaluated under semantically mismatched captions. In such cases, models frequently exhibit approximation behavior, producing a plausible bounding box that satisfies only part of the expression (\eg, preserving the original object while ignoring modified contextual cues). Because mismatched captions represent realistic edge cases, this behavior compromises reliability and raises concerns from an explainability perspective. Identifying its underlying causes is thus essential for improving model faithfulness and interpretability. Adopting a mechanistic interpretability viewpoint, this work examines whether embedding anisotropy contributes to counterfactual failures. A similarity-controlled counterfactual caption generation protocol is introduced to systematically perturb object or contextual components within predefined embedding similarity intervals, enabling a fine-grained analysis of grounding behavior as a function of alignment. Experiments on two Transformer-based models with markedly different embedding geometries (BERT-based TransVG and CLIP-based SwimVG) reveal no meaningful correlation between cosine similarity and approximation. These findings suggest that anisotropy alone does not account for counterfactual errors, and that robustness requires investigating finer-grained geometric properties of the embedding space.

Backdoor attacks on language models pose a significant threat to AI safety, where models behave normally on most inputs but exhibit harmful behavior when triggered by specific patterns. Detecting such backdoors through mechanistic interpretability remains an open challenge. We investigate two sparse autoencoder architectures -- Crosscoders and Differential SAEs (Diff-SAE) -- for isolating backdoor-related features in fine-tuned models. Using a controlled SQL injection backdoor triggered by year-based context ("2024" triggers vulnerable code, "2023" triggers safe code), we evaluate both approaches across LoRA and full-rank fine-tuning regimes on SmolLM2-360M. We find that Diff-SAE consistently and substantially outperforms Crosscoders for backdoor isolation. Diff-SAE achieves a Backdoor Isolation Score (BIS) of 0.40 with perfect precision (1.0) and zero false positive rate across most experimental conditions, while Crosscoders fail almost entirely with BIS below 0.02 in most cases. This performance gap holds across multiple transformer layers (14, 18, 22, 26) and both fine-tuning regimes, with full-rank fine-tuning producing particularly clean backdoor signals. Our results suggest that backdoors manifest as directional activation shifts rather than sparse feature activations, making difference-based representations fundamentally more effective for detection. These findings have important implications for AI safety monitoring and the development of interpretability tools for detecting model manipulation.

Automated crack inspection is increasingly recognized as a critical component of infrastructure monitoring; however, cracks continue to be reported primarily as binary segmentation masks by many current vision-based systems. While localization is facilitated by such masks, limited structural information is provided for robust engineering interpretation. For practical crack assessment, measurable morphological features -- including centerline geometry, branching behavior, junction locations, topology, and severity-related indicators -- are required. In this work, \textit{CrackMorph-XAI-Net}, an explainable morphology-aware framework for image-based crack analysis, is presented. Crack image and region-mask data are converted into a sequence of interpretable structural outputs through four distinct stages: topology-preserving skeleton extraction, junction detection via Gaussian heatmap regression, morphology descriptor computation, and severity-oriented screening. To support rigorous stage-wise evaluation, the standard \textit{CRACK500} benchmark is extended with aligned skeleton maps, junction heatmaps, and topology labels. Experimental validation demonstrates that a mean Dice coefficient of 0.991 is achieved by the learned skeleton extraction stage, with topology preserved in 98.5\% of test images. Furthermore, a recall of 0.964 and an F1-score of 0.887 are obtained in the junction detection stage, highlighting the efficacy of heatmap regression for sparse structural targets. Strong agreement between predicted and reference morphology values is revealed by descriptor-level evaluation, with correlations exceeding 0.95 for length, width, orientation, junction count, and tortuosity.