Grad-CAM for Vision Transformers: A Systematic Taxonomy and Audit of Methodological Ambiguity in Explainable AI

Gradient-weighted Class Activation Mapping (Grad-CAM) is widely used to visualize model decisions, but it was originally formulated for convolutional neural networks, where spatial feature maps and channel dimensions have clear architectural meanings. Vision Transformers (ViTs) do not provide the same structure, instead representing images through tokens, attention, residual streams, and multimodal interactions. This paper presents a systematic taxonomy and literature audit of how Grad-CAM and related methods are adapted, justified, and reported for ViT-based architectures. From an initial search of more than 550 papers, we identify 175 papers that apply Grad-CAM or Grad-CAM-adjacent methods to ViTs. We find that most papers do not provide a full mathematical or implementation-level account of how Grad-CAM is adapted to transformer representations. To characterize this gap, we introduce a descriptive taxonomy of ViT Grad-CAM adaptations that makes explicit the feature locations, gradient targets, spatial reconstruction steps, and aggregation choices that are often left implicit. This taxonomy is not intended to prescribe a single correct adaptation, but to clarify the range of methodological choices being made. The study shows that Grad-CAM on ViTs is often treated as a trivial extension of CNN-based Grad-CAM, despite requiring nontrivial choices that affect rigor, reproducibility, and interpretation.

Mechanistic Interpretability of LLM Jailbreaks via Internal Attribution Graphs

Large language models (LLMs) exhibit remarkable capabilities but remain highly vulnerable to adversarial prompts and jailbreak attacks. Existing approaches primarily analyze these failures through input-output behaviors or attribution methods, offering limited insight into how adversarial perturbations alter the model's internal reasoning. Consequently, the mechanisms underlying unsafe or incorrect behaviors remain poorly understood. We introduce a mechanistic framework for diagnosing LLM vulnerabilities using paired internal computation graphs, which represent prompt-specific inference as structured causal interactions among latent features. By constructing and aligning computation graphs for clean and attacked prompts, we reveal that adversarial attacks induce systematic transformations of internal reasoning, including suppression of safety-relevant components, emergence of attack-specific features, and rerouting of computation paths. Building on this representation, we propose a unified framework that (i) decomposes computation into invariant, suppressed, and emergent structures, (ii) identifies recurring vulnerability motifs associated with failure modes, and (iii) performs causal interventions on nodes, paths, and subgraphs to directly evaluate their contributions to attack success. This enables a transition from descriptive attribution to causal diagnosis of model failures. Experiments across multiple open-source LLMs and diverse adversarial and jailbreak benchmarks demonstrate that structural deviations in internal computation graphs strongly correlate with unsafe behaviors. Furthermore, targeted interventions on identified vulnerability motifs improve model robustness, establishing internal computation graphs as a principled foundation for understanding, diagnosing, and mitigating LLM vulnerabilities.

Explainable Novel Category Discovery in Semantic Concept Space

Novel category discovery aims to identify unseen classes from unlabeled data by transferring knowledge from labeled categories, but most existing methods perform discovery in opaque latent feature spaces. As a result, they may separate novel categories accurately while providing little insight into what semantic evidence defines each discovered group. We propose xNCD, an explainable novel category discovery framework that performs both representation-based discovery and pseudo-label assignment directly in a structured semantic concept space. Instead of clustering arbitrary deep features, xNCD learns a label-free concept representation by aligning visual features with vision-language similarity priors from pretrained multimodal models, and then applies a unified labeled-and-unlabeled self-labeling objective over concept-space logits. This design makes each discovered category explainable by construction through stable concept signatures and instance-level concept evidence. Theoretically, we show that routing discovery through a semantic concept bottleneck induces a strict restriction of the feature-space hypothesis class, excluding a large family of unconstrained decision rules and biasing induced partitions toward semantically interpretable concept coordinates. Experiments on CIFAR-10, CIFAR-100, and CUB-200 demonstrate that xNCD preserves strong discovery performance while providing intrinsic explanations. Under task-agnostic evaluation, xNCD achieves 92.63% overall accuracy on CIFAR-10, close to UNO's 93.4%, and improves CIFAR-100 overall accuracy from 73.2% to 76.45%, while being the only compared method that provides human-readable cluster- and instance-level explanations.

Robust and Clinically Reliable EEG Biomarkers: A Cross Population Framework for Generalizable Parkinson's Disease Detection

Developing robust and clinically reliable EEG biomarkers requires evaluation frameworks that explicitly address cross population generalization in multi site settings such as Parkinsons disease (PD) detection. Models trained under i.i.d. assumptions often capture population specific artifacts rather than disease relevant neural structure, leading to poor generalization across clinical cohorts. EEG further amplifies this challenge due to low signal to noise ratio and heterogeneous acquisition conditions. We propose a population aware evaluation framework to assess the robustness and clinical reliability of EEG biomarkers under distribution shift. Using an n gram expansion strategy, we enumerate all cross population train test configurations across five independent cohorts, resulting in 75 directional evaluations. A nested cross validation design with integrated channel selection ensures prospective biomarker identification without population leakage. Results show that cross population transfer is asymmetric and that both accuracy and biomarker stability improve with increasing training population diversity, achieving up to 94.1% accuracy on held out cohorts. A theoretical analysis based on mixture risk optimization and hypothesis space contraction explains these trends, showing that multi population training promotes population robust representations. This work establishes a principled framework for learning robust, generalizable, and clinically reliable EEG biomarkers for multi site biomedical applications.

Learning to Select Like Humans: Explainable Active Learning for Medical Imaging

Medical image analysis requires substantial labeled data for model training, yet expert annotation is expensive and time-consuming. Active learning (AL) addresses this challenge by strategically selecting the most informative samples for the annotation purpose, but traditional methods solely rely on predictive uncertainty while ignoring whether models learn from clinically meaningful features a critical requirement for clinical deployment. We propose an explainability-guided active learning framework that integrates spatial attention alignment into a sample acquisition process. Our approach advocates for a dual-criterion selection strategy combining: (i) classification uncertainty to identify informative examples, and (ii) attention misalignment with radiologist-defined regions-of-interest (ROIs) to target samples where the model focuses on incorrect features. By measuring misalignment between Grad-CAM attention maps and expert annotations using \emph{Dice similarity}, our acquisition function judiciously identifies samples that enhance both predictive performance and spatial interpretability. We evaluate the framework using three expert-annotated medical imaging datasets, namely, BraTS (MRI brain tumors), VinDr-CXR (chest X-rays), and SIIM-COVID-19 (chest X-rays). Using only 570 strategically selected samples, our explainability-guided approach consistently outperforms random sampling across all the datasets, achieving 77.22\% accuracy on BraTS, 52.37\% on VinDr-CXR, and 52.66\% on SIIM-COVID. Grad-CAM visualizations confirm that the models trained by our dual-criterion selection focus on diagnostically relevant regions, demonstrating that incorporating explanation guidance into sample acquisition yields superior data efficiency while maintaining clinical interpretability.

Acting Flatterers via LLMs Sycophancy: Combating Clickbait with LLMs Opposing-Stance Reasoning

The widespread proliferation of online content has intensified concerns about clickbait, deceptive or exaggerated headlines designed to attract attention. While Large Language Models (LLMs) offer a promising avenue for addressing this issue, their effectiveness is often hindered by Sycophancy, a tendency to produce reasoning that matches users' beliefs over truthful ones, which deviates from instruction-following principles. Rather than treating sycophancy as a flaw to be eliminated, this work proposes a novel approach that initially harnesses this behavior to generate contrastive reasoning from opposing perspectives. Specifically, we design a Self-renewal Opposing-stance Reasoning Generation (SORG) framework that prompts LLMs to produce high-quality agree and disagree reasoning pairs for a given news title without requiring ground-truth labels. To utilize the generated reasoning, we develop a local Opposing Reasoning-based Clickbait Detection (ORCD) model that integrates three BERT encoders to represent the title and its associated reasoning. The model leverages contrastive learning, guided by soft labels derived from LLM-generated credibility scores, to enhance detection robustness. Experimental evaluations on three benchmark datasets demonstrate that our method consistently outperforms LLM prompting, fine-tuned smaller language models, and state-of-the-art clickbait detection baselines.

Expert-Guided Explainable Few-Shot Learning with Active Sample Selection for Medical Image Analysis

Medical image analysis faces two critical challenges: scarcity of labeled data and lack of model interpretability, both hindering clinical AI deployment. Few-shot learning (FSL) addresses data limitations but lacks transparency in predictions. Active learning (AL) methods optimize data acquisition but overlook interpretability of acquired samples. We propose a dual-framework solution: Expert-Guided Explainable Few-Shot Learning (EGxFSL) and Explainability-Guided AL (xGAL). EGxFSL integrates radiologist-defined regions-of-interest as spatial supervision via Grad-CAM-based Dice loss, jointly optimized with prototypical classification for interpretable few-shot learning. xGAL introduces iterative sample acquisition prioritizing both predictive uncertainty and attention misalignment, creating a closed-loop framework where explainability guides training and sample selection synergistically. On the BraTS (MRI), VinDr-CXR (chest X-ray), and SIIM-COVID-19 (chest X-ray) datasets, we achieve accuracies of 92\%, 76\%, and 62\%, respectively, consistently outperforming non-guided baselines across all datasets. Under severe data constraints, xGAL achieves 76\% accuracy with only 680 samples versus 57\% for random sampling. Grad-CAM visualizations demonstrate guided models focus on diagnostically relevant regions, with generalization validated on breast ultrasound confirming cross-modality applicability.

Explainability-Guided Defense: Attribution-Aware Model Refinement Against Adversarial Data Attacks

The growing reliance on deep learning models in safety-critical domains such as healthcare and autonomous navigation underscores the need for defenses that are both robust to adversarial perturbations and transparent in their decision-making. In this paper, we identify a connection between interpretability and robustness that can be directly leveraged during training. Specifically, we observe that spurious, unstable, or semantically irrelevant features identified through Local Interpretable Model-Agnostic Explanations (LIME) contribute disproportionately to adversarial vulnerability. Building on this insight, we introduce an attribution-guided refinement framework that transforms LIME from a passive diagnostic into an active training signal. Our method systematically suppresses spurious features using feature masking, sensitivity-aware regularization, and adversarial augmentation in a closed-loop refinement pipeline. This approach does not require additional datasets or model architectures and integrates seamlessly into standard adversarial training. Theoretically, we derive an attribution-aware lower bound on adversarial distortion that formalizes the link between explanation alignment and robustness. Empirical evaluations on CIFAR-10, CIFAR-10-C, and CIFAR-100 demonstrate substantial improvements in adversarial robustness and out-of-distribution generalization.

CoSwin: Convolution Enhanced Hierarchical Shifted Window Attention For Small-Scale Vision

Vision Transformers (ViTs) have achieved impressive results in computer vision by leveraging self-attention to model long-range dependencies. However, their emphasis on global context often comes at the expense of local feature extraction in small datasets, particularly due to the lack of key inductive biases such as locality and translation equivariance. To mitigate this, we propose CoSwin, a novel feature-fusion architecture that augments the hierarchical shifted window attention with localized convolutional feature learning. Specifically, CoSwin integrates a learnable local feature enhancement module into each attention block, enabling the model to simultaneously capture fine-grained spatial details and global semantic structure. We evaluate CoSwin on multiple image classification benchmarks including CIFAR-10, CIFAR-100, MNIST, SVHN, and Tiny ImageNet. Our experimental results show consistent performance gains over state-of-the-art convolutional and transformer-based models. Notably, CoSwin achieves improvements of 2.17% on CIFAR-10, 4.92% on CIFAR-100, 0.10% on MNIST, 0.26% on SVHN, and 4.47% on Tiny ImageNet over the baseline Swin Transformer. These improvements underscore the effectiveness of local-global feature fusion in enhancing the generalization and robustness of transformers for small-scale vision. Code and pretrained weights available at https://github.com/puskal-khadka/coswin

Ecologically Valid Benchmarking and Adaptive Attention: Scalable Marine Bioacoustic Monitoring

Underwater Passive Acoustic Monitoring (UPAM) provides rich spatiotemporal data for long-term ecological analysis, but intrinsic noise and complex signal dependencies hinder model stability and generalization. Multilayered windowing has improved target sound localization, yet variability from shifting ambient noise, diverse propagation effects, and mixed biological and anthropogenic sources demands robust architectures and rigorous evaluation. We introduce GetNetUPAM, a hierarchical nested cross-validation framework designed to quantify model stability under ecologically realistic variability. Data are partitioned into distinct site-year segments, preserving recording heterogeneity and ensuring each validation fold reflects a unique environmental subset, reducing overfitting to localized noise and sensor artifacts. Site-year blocking enforces evaluation against genuine environmental diversity, while standard cross-validation on random subsets measures generalization across UPAM's full signal distribution, a dimension absent from current benchmarks. Using GetNetUPAM as the evaluation backbone, we propose the Adaptive Resolution Pooling and Attention Network (ARPA-N), a neural architecture for irregular spectrogram dimensions. Adaptive pooling with spatial attention extends the receptive field, capturing global context without excessive parameters. Under GetNetUPAM, ARPA-N achieves a 14.4% gain in average precision over DenseNet baselines and a log2-scale order-of-magnitude drop in variability across all metrics, enabling consistent detection across site-year folds and advancing scalable, accurate bioacoustic monitoring.