Forget and Rewire (FaR) methodology has demonstrated strong resilience against Bit-Flip Attacks (BFAs) on Transformer-based models by obfuscating critical parameters through dynamic rewiring of linear layers. However, the application of FaR introduces non-negligible performance and memory overheads, primarily due to the runtime modification of activation pathways and the lack of hardware-level optimization. To overcome these limitations, we propose FaRAccel, a novel hardware accelerator architecture implemented on FPGA, specifically designed to offload and optimize FaR operations. FaRAccel integrates reconfigurable logic for dynamic activation rerouting, and lightweight storage of rewiring configurations, enabling low-latency inference with minimal energy overhead. We evaluate FaRAccel across a suite of Transformer models and demonstrate substantial reductions in FaR inference latency and improvement in energy efficiency, while maintaining the robustness gains of the original FaR methodology. To the best of our knowledge, this is the first hardware-accelerated defense against BFAs in Transformers, effectively bridging the gap between algorithmic resilience and efficient deployment on real-world AI platforms.

Pressure ulcers (PUs) are a serious and prevalent healthcare concern. Accurate classification of PU severity (Stages I-IV) is essential for proper treatment but remains challenging due to subtle visual distinctions and subjective interpretation, leading to variability among clinicians. Prior AI-based approaches using Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) achieved promising accuracy but offered limited interpretability. We present FT-ARM (Fine-Tuned Agentic Reflection Multimodal model), a fine-tuned multimodal large language model (MLLM) with an agentic self-reflection mechanism for pressure ulcer severity classification. Inspired by clinician-style diagnostic reassessment, FT-ARM iteratively refines its predictions by reasoning over visual features and encoded clinical knowledge from text, enhancing both accuracy and consistency. On the publicly available Pressure Injury Image Dataset (PIID), FT-ARM, fine-tuned from LLaMA 3.2 90B, achieved 85% accuracy in classifying PU stages I-IV, surpassing prior CNN-based models by +4%. Unlike earlier CNN/ViT studies that relied solely on offline evaluations, FT-ARM is designed and tested for live inference, reflecting real-time deployment conditions. Furthermore, it produces clinically grounded natural-language explanations, improving interpretability and trust. By integrating fine-tuning and reflective reasoning across multimodal inputs, FT-ARM advances the reliability, transparency, and clinical applicability of automated wound assessment systems, addressing the critical need for consistent and explainable PU staging to support improved patient care.

Abstract--Vision Transformers (ViTs) have emerged as powerful architectures in medical image analysis, excelling in tasks such as disease detection, segmentation, and classification. However, their reliance on large, attention-driven models makes them vulnerable to hardware-level attacks. In this paper, we propose a novel threat model referred to as Med-Hammer that combines the Rowhammer hardware fault injection with neural Trojan attacks to compromise the integrity of ViT -based medical imaging systems. Specifically, we demonstrate how malicious bit flips induced via Rowhammer can trigger implanted neural Trojans, leading to targeted misclassification or suppression of critical diagnoses (e.g., tumors or lesions) in medical scans. Through extensive experiments on benchmark medical imaging datasets such as ISIC, Brain T umor, and MedMNIST, we show that such attacks can remain stealthy while achieving high attack success rates about 82.51% and 92.56% in MobileViT and SwinTrans-former, respectively. We further investigate how architectural properties, such as model sparsity, attention weight distribution, and number of features of the layer, impact attack effectiveness. Our findings highlight a critical and underexplored intersection between hardware-level faults and deep learning security in healthcare applications, underscoring the urgent need for robust defenses spanning both model architectures and underlying hardware platforms. In clinical practice, medical imaging plays a central role in detecting, diagnosing, and monitoring a wide range of conditions.

We show that across architecture (Transformer vs. Mamba vs. RWKV), training dataset (OpenWebText vs. The Pile), and scale (14 million parameters to 12 billion parameters), autoregressive language models exhibit highly consistent patterns of change in their behavior over the course of pretraining. Based on our analysis of over 1,400 language model checkpoints on over 110,000 tokens of English, we find that up to 98% of the variance in language model behavior at the word level can be explained by three simple heuristics: the unigram probability (frequency) of a given word, the $n$-gram probability of the word, and the semantic similarity between the word and its context. Furthermore, we see consistent behavioral phases in all language models, with their predicted probabilities for words overfitting to those words' $n$-gram probabilities for increasing $n$ over the course of training. Taken together, these results suggest that learning in neural language models may follow a similar trajectory irrespective of model details.

Assessing scenario coverage is crucial for evaluating the robustness of autonomous agents, yet existing methods rely on expensive human annotations or computationally intensive Large Vision-Language Models (LVLMs). These approaches are impractical for large-scale deployment due to cost and efficiency constraints. To address these shortcomings, we propose SCOUT (Scenario Coverage Oversight and Understanding Tool), a lightweight surrogate model designed to predict scenario coverage labels directly from an agent's latent sensor representations. SCOUT is trained through a distillation process, learning to approximate LVLM-generated coverage labels while eliminating the need for continuous LVLM inference or human annotation. By leveraging precomputed perception features, SCOUT avoids redundant computations and enables fast, scalable scenario coverage estimation. We evaluate our method across a large dataset of real-life autonomous navigation scenarios, demonstrating that it maintains high accuracy while significantly reducing computational cost. Our results show that SCOUT provides an effective and practical alternative for large-scale coverage analysis. While its performance depends on the quality of LVLM-generated training labels, SCOUT represents a major step toward efficient scenario coverage oversight in autonomous systems.

Despite their impressive performance, vision-language models (VLMs) still struggle on culturally situated inputs. To understand how VLMs process culturally grounded information, we study the presence of culture-sensitive neurons, i.e. neurons whose activations show preferential sensitivity to inputs associated with particular cultural contexts. We examine whether such neurons are important for culturally diverse visual question answering and where they are located. Using the CVQA benchmark, we identify neurons of culture selectivity and perform causal tests by deactivating the neurons flagged by different identification methods. Experiments on three VLMs across 25 cultural groups demonstrate the existence of neurons whose ablation disproportionately harms performance on questions about the corresponding cultures, while having minimal effects on others. Moreover, we propose a new margin-based selector - Contrastive Activation Selection (CAS), and show that it outperforms existing probability- and entropy-based methods in identifying culture-sensitive neurons. Finally, our layer-wise analyses reveals that such neurons tend to cluster in certain decoder layers. Overall, our findings shed new light on the internal organization of multimodal representations.

Large Language Models have demonstrated strong performance on many established reasoning benchmarks. However, these benchmarks primarily evaluate structured skills like quantitative problem-solving, leaving a gap in assessing flexible, multifaceted reasoning abilities that are central to human intelligence. These abilities require integrating logical deduction with spatial awareness and constraint satisfaction, which current evaluations do not measure well. To address this, we introduce RiddleBench, a benchmark of 1,737 challenging puzzles in English designed to probe these core reasoning capabilities. Evaluation of state-of-the-art models on RiddleBench shows fundamental weaknesses. Even top proprietary models like Gemini 2.5 Pro, o3, and Claude 4 Sonnet achieve accuracy just above 60% (60.30%, 63.37%, and 63.16%). Analysis further reveals deep failures, including hallucination cascades (accepting flawed reasoning from other models) and poor self-correction due to a strong self-confirmation bias. Their reasoning is also fragile, with performance degrading significantly when constraints are reordered or irrelevant information is introduced. RiddleBench functions as a diagnostic tool for these issues and as a resource for guiding the development of more robust and reliable language models.

In multimodal learning, dominant modalities often overshadow others, limiting generalization. We propose Modality-Aware Sharpness-Aware Minimization (M-SAM), a model-agnostic framework that applies to many modalities and supports early and late fusion scenarios. In every iteration, M-SAM in three steps optimizes learning. \textbf{First, it identifies the dominant modality} based on modalities' contribution in the accuracy using Shapley. \textbf{Second, it decomposes the loss landscape}, or in another language, it modulates the loss to prioritize the robustness of the model in favor of the dominant modality, and \textbf{third, M-SAM updates the weights} by backpropagation of modulated gradients. This ensures robust learning for the dominant modality while enhancing contributions from others, allowing the model to explore and exploit complementary features that strengthen overall performance. Extensive experiments on four diverse datasets show that M-SAM outperforms the latest state-of-the-art optimization and gradient manipulation methods and significantly balances and improves multimodal learning.

The growing integration of artificial intelligence (AI) into human cognition raises a fundamental question: does AI merely improve efficiency, or does it alter how we think? This study experimentally tested whether short-term exposure to narrow AI tools enhances core cognitive abilities or simply optimizes task performance. Thirty young adults completed standardized neuropsychological assessments embedded in a seven-week protocol with a four-week online intervention involving problem-solving and verbal comprehension tasks, either with or without AI support (ChatGPT). While AI-assisted participants completed several tasks faster and more accurately, no significant pre-post differences emerged in standardized measures of problem solving or verbal comprehension. These results demonstrate efficiency gains without cognitive change, suggesting that current narrow AI systems serve as cognitive scaffolds extending performance without transforming underlying mental capacities. The findings highlight the need for ethical and educational frameworks that promote critical and autonomous thinking in an increasingly AI-augmented cognitive ecology.

Large language models (LLMs) have demonstrated significant potential to accelerate scientific discovery as valuable tools for analyzing data, generating hypotheses, and supporting innovative approaches in various scientific fields. In this work, we investigate how LLMs can handle the transition from conceptual research ideas to well-structured research plans. Effective research planning not only supports scientists in advancing their research but also represents a crucial capability for the development of autonomous research agents. Despite its importance, the field lacks a systematic understanding of LLMs' research planning capability. To rigorously measure this capability, we introduce the Idea2Plan task and Idea2Plan Bench, a benchmark built from 200 ICML 2025 Spotlight and Oral papers released after major LLM training cutoffs. Each benchmark instance includes a research idea and a grading rubric capturing the key components of valid plans. We further propose Idea2Plan JudgeEval, a complementary benchmark to assess the reliability of LLM-based judges against expert annotations. Experimental results show that GPT-5 and GPT-5-mini achieve the strongest performance on the benchmark, though substantial headroom remains for future improvement. Our study provides new insights into LLMs' capability for research planning and lay the groundwork for future progress.