A Benchmark for Hallucination Detection in VLMs for Gastrointestinal Endoscopy

Vision-language models (VLMs) are prone to hallucination, which remains a major barrier to their safe deployment in clinical practice. To date, most hallucination detection methods have been evaluated on radiology benchmarks such as MIMIC-CXR and VQA-RAD, while gastrointestinal (GI) endoscopy remains largely underexplored. In this paper, we benchmark nine hallucination detection methods on the Gut-VLM dataset, a GI diagnostic Visual Question Answering (VQA) dataset with 4,392 test VQA pairs, across five VLMs (MedGemma-4B, MedGemma-27B, LLaVA-Med-7B, LLaVA-v1.6-7B, and Lingshu-32B). The methods span three categories: black-box methods (RadFlag, SelfCheckGPT-NLI), gray-box methods (AvgProb, AvgEnt, MaxProb, MaxEnt, Semantic Entropy, and VASE), and a white-box method (ReXTrust). Our results show that ReXTrust, a white-box method, achieves the highest AUC across all five models, outperforming the strongest alternative method on each VLM by a statistically significant margin (paired permutation test, p < 0.001 in all cases), reaching a peak AUC of 93.0 on MedGemma-4B. White-box hidden-state access provides a consistent advantage of 19.5 AUC points on average (range: 9.5--33.5), with ReXTrust maintaining strong performance even on LLaVA-v1.6-7B (AUC 79.9), where black-box methods and clustering-based gray-box methods collapse to near-chance performance. Among non-white-box methods, token-level gray-box statistics (MaxEnt, MaxProb) are the strongest alternatives, outperforming both clustering-based gray-box methods (Semantic Entropy, VASE) and black-box approaches on average. We further identify confident confabulation, a failure mode in which models hallucinate with high inter-sample consistency or high token-level probability, as a systemic failure for both consistency and uncertainty-based methods.

When Confidence Lacks Concepts: Interpretable OOD Detection via Representation Perturbations

Deep neural networks have achieved remarkable performance across medical imaging tasks, yet their tendency to overgeneralize under distributional shifts poses a major obstacle to safe clinical deployment. Out-of-Distribution (OOD) detection methods aim to mitigate this risk, but most existing approaches rely on opaque internal signals with poorly understood semantic meaning, limiting trust in safety-critical settings. In this work, we propose an interpretable OOD detection framework that probes the stability of model predictions under class-conditioned semantic perturbations. Leveraging sparse autoencoders (SAEs), we learn class-specific concept vectors from in-distribution data that disentangle dense intermediate representations into sparse, semantically meaningful components. At inference, we perturb deeper-layer representations using the concept vectors associated with the model's predicted class and measure the class logits stability. We hypothesize that in-distribution samples exhibit low sensitivity to such perturbations, as their representations align with class-specific semantic directions, whereas OOD samples show amplified deviations due to representational misalignment. By framing OOD detection as a concept conditioned stability analysis, our approach provides both a discriminative OOD signal and an interpretable lens into the internal mechanisms driving model uncertainty, making it particularly suitable for high stakes medical applications.

BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation

High-Dimensional Low-Sample Size (HDLSS) tabular domains (e.g., omics) are characterized by $n \ll m$, where $n$ = number of samples, and $m$ = number of features. Such domains often exhibit strong local correlation groups, sparse cross-group dependencies, heavy-tailed non-Gaussian marginals, heteroscedastic noise, and structured missingness, making direct density learning in $\mathbb{R}^m$ ill-conditioned since $n \ll m$. We propose BSTabDiff, a block-subunit generative framework that partitions the $m$ observed features into $M$ latent blocks ($M \ll m$) and generates each block via a shared low-dimensional subunit variable, concentrating global dependence learning in the compact block-latent space $\mathbb{R}^M$ while decoding to the full feature space with copula-driven dependence, flexible per-feature marginals, and explicit missingness mechanisms. BSTabDiff supports modern deep priors on block latents, including diffusion and normalizing flows, enabling stable synthesis and controllable benchmark generation in the HDLSS regime. Empirically, BSTabDiff produces more realistic and stable high-dimensional synthetic data when compared with unstructured tabular generators on HDLSS data.

Two is better than one: A Collapse-free Multi-Reward RLIF Training Framework

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning ability of LLMs, but often depends on external supervision from human annotations or gold-standard solutions. Reinforcement learning from internal feedback (RLIF) has recently emerged as a scalable unsupervised alternative, using signals extracted from the model itself. However, existing RLIF methods typically rely on a single internal reward, which can lead to reward hacking, entropy collapse, and degraded reasoning structure. We propose a multi-reward RLIF framework that decomposes the training signal into two complementary components: an answer-level reward based on cluster voting and a completion-level reward based on token-wise self-certainty. To combine these signals robustly, we apply GDPO-based normalization to reduce reward-scale imbalance. We further introduce KL-Cov regularization, which targets low-entropy token distributions responsible for disproportionate entropy reduction, preserving exploration and preventing late-stage collapse. Across mathematical reasoning and code-generation benchmarks, our method improves stability and robustness over prior unsupervised RL approaches, while achieving performance close to supervised RLVR methods. These results show that complementary internal rewards, combined with targeted regularization, can support stable long-horizon reasoning without relying on external ground-truth supervision. Code will be released soon.

Local K-Similarity Constraint for Federated Learning with Label Noise

Federated learning on clients with noisy labels is a challenging problem, as such clients can infiltrate the global model, impacting the overall generalizability of the system. Existing methods proposed to handle noisy clients assume that a sufficient number of clients with clean labels are available, which can be leveraged to learn a robust global model while dampening the impact of noisy clients. This assumption fails when a high number of heterogeneous clients contain noisy labels, making the existing approaches ineffective. In such scenarios, it is important to locally regularize the clients before communication with the global model, to ensure the global model isn't corrupted by noisy clients. While pre-trained self-supervised models can be effective for local regularization, existing centralized approaches relying on pretrained initialization are impractical in a federated setting due to the potentially large size of these models, which increases communication costs. In that line, we propose a regularization objective for client models that decouples the pre-trained and classification models by enforcing similarity between close data points within the client. We leverage the representation space of a self-supervised pretrained model to evaluate the closeness among examples. This regularization, when applied with the standard objective function for the downstream task in standard noisy federated settings, significantly improves performance, outperforming existing state-of-the-art federated methods in multiple computer vision and medical image classification benchmarks. Unlike other techniques that rely on self-supervised pretrained initialization, our method does not require the pretrained model and classifier backbone to share the same architecture, making it architecture-agnostic.

Effect of Data Augmentation on Conformal Prediction for Diabetic Retinopathy

The clinical deployment of deep learning models for high-stakes tasks such as diabetic retinopathy (DR) grading requires demonstrable reliability. While models achieve high accuracy, their clinical utility is limited by a lack of robust uncertainty quantification. Conformal prediction (CP) offers a distribution-free framework to generate prediction sets with statistical guarantees of coverage. However, the interaction between standard training practices like data augmentation and the validity of these guarantees is not well understood. In this study, we systematically investigate how different data augmentation strategies affect the performance of conformal predictors for DR grading. Using the DDR dataset, we evaluate two backbone architectures -- ResNet-50 and a Co-Scale Conv-Attentional Transformer (CoaT) -- trained under five augmentation regimes: no augmentation, standard geometric transforms, CLAHE, Mixup, and CutMix. We analyze the downstream effects on conformal metrics, including empirical coverage, average prediction set size, and correct efficiency. Our results demonstrate that sample-mixing strategies like Mixup and CutMix not only improve predictive accuracy but also yield more reliable and efficient uncertainty estimates. Conversely, methods like CLAHE can negatively impact model certainty. These findings highlight the need to co-design augmentation strategies with downstream uncertainty quantification in mind to build genuinely trustworthy AI systems for medical imaging.

Addressing Bias in VLMs for Glaucoma Detection Without Protected Attribute Supervision

Vision-Language Models (VLMs) have achieved remarkable success on multimodal tasks such as image-text retrieval and zero-shot classification, yet they can exhibit demographic biases even when explicit protected attributes are absent during training. In this work, we focus on automated glaucoma screening from retinal fundus images, a critical application given that glaucoma is a leading cause of irreversible blindness and disproportionately affects underserved populations. Building on a reweighting-based contrastive learning framework, we introduce an attribute-agnostic debiasing method that (i) infers proxy subgroups via unsupervised clustering of image-image embeddings, (ii) computes gradient-similarity weights between the CLIP-style multimodal loss and a SimCLR-style image-pair contrastive loss, and (iii) applies these weights in a joint, top-$k$ weighted objective to upweight underperforming clusters. This label-free approach adaptively targets the hardest examples, thereby reducing subgroup disparities. We evaluate our method on the Harvard FairVLMed glaucoma subset, reporting Equalized Odds Distance (EOD), Equalized Subgroup AUC (ES AUC), and Groupwise AUC to demonstrate equitable performance across inferred demographic subgroups.

NERO: Explainable Out-of-Distribution Detection with Neuron-level Relevance

Ensuring reliability is paramount in deep learning, particularly within the domain of medical imaging, where diagnostic decisions often hinge on model outputs. The capacity to separate out-of-distribution (OOD) samples has proven to be a valuable indicator of a model's reliability in research. In medical imaging, this is especially critical, as identifying OOD inputs can help flag potential anomalies that might otherwise go undetected. While many OOD detection methods rely on feature or logit space representations, recent works suggest these approaches may not fully capture OOD diversity. To address this, we propose a novel OOD scoring mechanism, called NERO, that leverages neuron-level relevance at the feature layer. Specifically, we cluster neuron-level relevance for each in-distribution (ID) class to form representative centroids and introduce a relevance distance metric to quantify a new sample's deviation from these centroids, enhancing OOD separability. Additionally, we refine performance by incorporating scaled relevance in the bias term and combining feature norms. Our framework also enables explainable OOD detection. We validate its effectiveness across multiple deep learning architectures on the gastrointestinal imaging benchmarks Kvasir and GastroVision, achieving improvements over state-of-the-art OOD detection methods.

NCDD: Nearest Centroid Distance Deficit for Out-Of-Distribution Detection in Gastrointestinal Vision

The integration of deep learning tools in gastrointestinal vision holds the potential for significant advancements in diagnosis, treatment, and overall patient care. A major challenge, however, is these tools' tendency to make overconfident predictions, even when encountering unseen or newly emerging disease patterns, undermining their reliability. We address this critical issue of reliability by framing it as an out-of-distribution (OOD) detection problem, where previously unseen and emerging diseases are identified as OOD examples. However, gastrointestinal images pose a unique challenge due to the overlapping feature representations between in- Distribution (ID) and OOD examples. Existing approaches often overlook this characteristic, as they are primarily developed for natural image datasets, where feature distinctions are more apparent. Despite the overlap, we hypothesize that the features of an in-distribution example will cluster closer to the centroids of their ground truth class, resulting in a shorter distance to the nearest centroid. In contrast, OOD examples maintain an equal distance from all class centroids. Based on this observation, we propose a novel nearest-centroid distance deficit (NCCD) score in the feature space for gastrointestinal OOD detection. Evaluations across multiple deep learning architectures and two publicly available benchmarks, Kvasir2 and Gastrovision, demonstrate the effectiveness of our approach compared to several state-of-the-art methods. The code and implementation details are publicly available at: https://github.com/bhattarailab/NCDD

Is Generative AI the Next Tactical Cyber Weapon For Threat Actors? Unforeseen Implications of AI Generated Cyber Attacks

In an era where digital threats are increasingly sophisticated, the intersection of Artificial Intelligence and cybersecurity presents both promising defenses and potent dangers. This paper delves into the escalating threat posed by the misuse of AI, specifically through the use of Large Language Models (LLMs). This study details various techniques like the switch method and character play method, which can be exploited by cybercriminals to generate and automate cyber attacks. Through a series of controlled experiments, the paper demonstrates how these models can be manipulated to bypass ethical and privacy safeguards to effectively generate cyber attacks such as social engineering, malicious code, payload generation, and spyware. By testing these AI generated attacks on live systems, the study assesses their effectiveness and the vulnerabilities they exploit, offering a practical perspective on the risks AI poses to critical infrastructure. We also introduce Occupy AI, a customized, finetuned LLM specifically engineered to automate and execute cyberattacks. This specialized AI driven tool is adept at crafting steps and generating executable code for a variety of cyber threats, including phishing, malware injection, and system exploitation. The results underscore the urgency for ethical AI practices, robust cybersecurity measures, and regulatory oversight to mitigate AI related threats. This paper aims to elevate awareness within the cybersecurity community about the evolving digital threat landscape, advocating for proactive defense strategies and responsible AI development to protect against emerging cyber threats.