Research Papers

The proliferation of Large Language Models (LLMs) has demonstrated remarkable capabilities, elevating the critical importance of LLM safety. However, existing safety methods rely on ad-hoc taxonomy and lack a rigorous, systematic protection, failing to ensure safety for the nuanced and complex behaviors of modern LLM systems. To address this problem, we solve LLM safety from legal compliance perspectives, named safety compliance. In this work, we posit relevant established legal frameworks as safety standards for defining and measuring safety compliance, including the EU AI Act and GDPR, which serve as core legal frameworks for AI safety and data security in Europe. To bridge the gap between LLM safety and legal compliance, we first develop a new benchmark for safety compliance by generating realistic LLM safety scenarios seeded with legal statutes. Subsequently, we align Qwen3-8B using Group Policy Optimization (GRPO) to construct a safety reasoner, Compliance Reasoner, which effectively aligns LLMs with legal standards to mitigate safety risks. Our comprehensive experiments demonstrate that the Compliance Reasoner achieves superior performance on the new benchmark, with average improvements of +10.45% for the EU AI Act and +11.85% for GDPR.

Recent advances in vision-language models (VLMs) have achieved remarkable performance on standard medical benchmarks, yet their true clinical reasoning ability remains unclear. Existing datasets predominantly emphasize classification accuracy, creating an evaluation illusion in which models appear proficient while still failing at high-stakes diagnostic reasoning. We introduce Neural-MedBench, a compact yet reasoning-intensive benchmark specifically designed to probe the limits of multimodal clinical reasoning in neurology. Neural-MedBench integrates multi-sequence MRI scans, structured electronic health records, and clinical notes, and encompasses three core task families: differential diagnosis, lesion recognition, and rationale generation. To ensure reliable evaluation, we develop a hybrid scoring pipeline that combines LLM-based graders, clinician validation, and semantic similarity metrics. Through systematic evaluation of state-of-the-art VLMs, including GPT-4o, Claude-4, and MedGemma, we observe a sharp performance drop compared to conventional datasets. Error analysis shows that reasoning failures, rather than perceptual errors, dominate model shortcomings. Our findings highlight the necessity of a Two-Axis Evaluation Framework: breadth-oriented large datasets for statistical generalization, and depth-oriented, compact benchmarks such as Neural-MedBench for reasoning fidelity. We release Neural-MedBench at https://neuromedbench.github.io/ as an open and extensible diagnostic testbed, which guides the expansion of future benchmarks and enables rigorous yet cost-effective assessment of clinically trustworthy AI.

On The Reproducibility Limitations of RAG Systems

Baiqiang Wang, Dongfang Zhao, Nathan R Tallent +1

Retrieval-Augmented Generation (RAG) is increasingly employed in generative AI-driven scientific workflows to integrate rapidly evolving scientific knowledge bases, yet its reliability is frequently compromised by non-determinism in their retrieval components. This paper introduces ReproRAG, a comprehensive benchmarking framework designed to systematically measure and quantify the reproducibility of vector-based retrieval systems. ReproRAG investigates sources of uncertainty across the entire pipeline, including different embedding models, precision, retrieval algorithms, hardware configurations, and distributed execution environments. Utilizing a suite of metrics, such as Exact Match Rate, Jaccard Similarity, and Kendall's Tau, the proposed framework effectively characterizes the trade-offs between reproducibility and performance. Our large-scale empirical study reveals critical insights; for instance, we observe that different embedding models have remarkable impact on RAG reproducibility. The open-sourced ReproRAG framework provides researchers and engineers productive tools to validate deployments, benchmark reproducibility, and make informed design decisions, thereby fostering more trustworthy AI for science.

Clinical machine learning faces a critical dilemma in high-stakes medical applications: algorithms achieving optimal diagnostic performance typically sacrifice the interpretability essential for physician decision-making, while interpretable methods compromise sensitivity in complex scenarios. This paradox becomes particularly acute in non-invasive prenatal testing (NIPT), where missed chromosomal abnormalities carry profound clinical consequences yet regulatory frameworks mandate explainable AI systems. We introduce Medical Priority Fusion (MPF), a constrained multi-objective optimization framework that resolves this fundamental trade-off by systematically integrating Naive Bayes probabilistic reasoning with Decision Tree rule-based logic through mathematically-principled weighted fusion under explicit medical constraints. Rigorous validation on 1,687 real-world NIPT samples characterized by extreme class imbalance (43.4:1 normal-to-abnormal ratio) employed stratified 5-fold cross-validation with comprehensive ablation studies and statistical hypothesis testing using McNemar's paired comparisons. MPF achieved simultaneous optimization of dual objectives: 89.3% sensitivity (95% CI: 83.9-94.7%) with 80% interpretability score, significantly outperforming individual algorithms (McNemar's test, p < 0.001). The optimal fusion configuration achieved Grade A clinical deployment criteria with large effect size (d = 1.24), establishing the first clinically-deployable solution that maintains both diagnostic accuracy and decision transparency essential for prenatal care. This work demonstrates that medical-constrained algorithm fusion can resolve the interpretability-performance trade-off, providing a mathematical framework for developing high-stakes medical decision support systems that meet both clinical efficacy and explainability requirements.

Early and accurate detection through Pap smear analysis is critical to improving patient outcomes and reducing mortality of Cervical cancer. State-of-the-art (SOTA) Convolutional Neural Networks (CNNs) require substantial computational resources, extended training time, and large datasets. In this study, a lightweight CNN model, S-Net (Simple Net), is developed specifically for cervical cancer detection and classification using Pap smear images to address these limitations. Alongside S-Net, six SOTA CNNs were evaluated using transfer learning, including multi-path (DenseNet201, ResNet152), depth-based (Serasnet152), width-based multi-connection (Xception), depth-wise separable convolutions (MobileNetV2), and spatial exploitation-based (VGG19). All models, including S-Net, achieved comparable accuracy, with S-Net reaching 99.99%. However, S-Net significantly outperforms the SOTA CNNs in terms of computational efficiency and inference time, making it a more practical choice for real-time and resource-constrained applications. A major limitation in CNN-based medical diagnosis remains the lack of transparency in the decision-making process. To address this, Explainable AI (XAI) techniques, such as SHAP, LIME, and Grad-CAM, were employed to visualize and interpret the key image regions influencing model predictions. The novelty of this study lies in the development of a highly accurate yet computationally lightweight model (S-Net) caPable of rapid inference while maintaining interpretability through XAI integration. Furthermore, this work analyzes the behavior of SOTA CNNs, investigates the effects of negative transfer learning on Pap smear images, and examines pixel intensity patterns in correctly and incorrectly classified samples.

AI-driven digital ecosystems span diverse stakeholders including technology firms, regulators, accelerators and civil society, yet often lack cohesive ethical governance. This paper proposes a four-pillar framework (SCOR) to embed accountability, fairness, and inclusivity across such multi-actor networks. Leveraging a design science approach, we develop a Shared Ethical Charter(S), structured Co-Design and Stakeholder Engagement protocols(C), a system of Continuous Oversight and Learning(O), and Adaptive Regulatory Alignment strategies(R). Each component includes practical guidance, from lite modules for resource-constrained start-ups to in-depth auditing systems for larger consortia. Through illustrative vignettes in healthcare, finance, and smart city contexts, we demonstrate how the framework can harmonize organizational culture, leadership incentives, and cross-jurisdictional compliance. Our mixed-method KPI design further ensures that quantitative targets are complemented by qualitative assessments of user trust and cultural change. By uniting ethical principles with scalable operational structures, this paper offers a replicable pathway toward responsible AI innovation in complex digital ecosystems.

The integration of artificial intelligence (AI) into telecommunications infrastructure introduces novel risks, such as algorithmic bias and unpredictable system behavior, that fall outside the scope of traditional cybersecurity and data protection frameworks. This paper introduces a precise definition and a detailed typology of telecommunications AI incidents, establishing them as a distinct category of risk that extends beyond conventional cybersecurity and data protection breaches. It argues for their recognition as a distinct regulatory concern. Using India as a case study for jurisdictions that lack a horizontal AI law, the paper analyzes the country's key digital regulations. The analysis reveals that India's existing legal instruments, including the Telecommunications Act, 2023, the CERT-In Rules, and the Digital Personal Data Protection Act, 2023, focus on cybersecurity and data breaches, creating a significant regulatory gap for AI-specific operational incidents, such as performance degradation and algorithmic bias. The paper also examines structural barriers to disclosure and the limitations of existing AI incident repositories. Based on these findings, the paper proposes targeted policy recommendations centered on integrating AI incident reporting into India's existing telecom governance. Key proposals include mandating reporting for high-risk AI failures, designating an existing government body as a nodal agency to manage incident data, and developing standardized reporting frameworks. These recommendations aim to enhance regulatory clarity and strengthen long-term resilience, offering a pragmatic and replicable blueprint for other nations seeking to govern AI risks within their existing sectoral frameworks.

Explaining the Reputational Risks of AI-Mediated Communication: Messages labeled as AI-assisted are viewed as less diagnostic of the sender's moral character

Pranav Khadpe, Kimi Wenzel, George Loewenstein +1

When someone sends us a thoughtful message, we naturally form judgments about their character. But what happens when that message carries a label indicating it was written with the help of AI? This paper investigates how the appearance of AI assistance affects our perceptions of message senders. Adding nuance to previous research, through two studies (N=399) featuring vignette scenarios, we find that AI-assistance labels don't necessarily make people view senders negatively. Rather, they dampen the strength of character signals in communication. We show that when someone sends a warmth-signalling message (like thanking or apologizing) without AI help, people more strongly categorize the sender as warm. At the same time, when someone sends a coldness-signalling message (like bragging or blaming) without assistance, people more confidently categorize them as cold. Interestingly, AI labels weaken both these associations: An AI-assisted apology makes the sender appear less warm than if they had written it themselves, and an AI-assisted blame makes the sender appear less cold than if they had composed it independently. This supports our signal diagnosticity explanation: messages labeled as AI-assisted are viewed as less diagnostic than messages which seem unassisted. We discuss how our findings shed light on the causal origins of previously reported observations in AI-Mediated Communication.

Personalized AI assistants often struggle to incorporate complex personal data and causal knowledge, leading to generic advice that lacks explanatory power. We propose REMI, a Causal Schema Memory architecture for a multimodal lifestyle agent that integrates a personal causal knowledge graph, a causal reasoning engine, and a schema based planning module. The idea is to deliver explainable, personalized recommendations in domains like fashion, personal wellness, and lifestyle planning. Our architecture uses a personal causal graph of the user's life events and habits, performs goal directed causal traversals enriched with external knowledge and hypothetical reasoning, and retrieves adaptable plan schemas to generate tailored action plans. A Large Language Model orchestrates these components, producing answers with transparent causal explanations. We outline the CSM system design and introduce new evaluation metrics for personalization and explainability, including Personalization Salience Score and Causal Reasoning Accuracy, to rigorously assess its performance. Results indicate that CSM based agents can provide more context aware, user aligned recommendations compared to baseline LLM agents. This work demonstrates a novel approach to memory augmented, causal reasoning in personalized agents, advancing the development of transparent and trustworthy AI lifestyle assistants.

Artificial intelligence deployment for automated photovoltaic (PV) monitoring faces interpretability barriers that limit adoption in energy infrastructure applications. While deep learning achieves high accuracy in thermal fault detection, validation that model decisions align with thermal physics principles remains lacking, creating deployment hesitancy where understanding model reasoning is critical. This study provides a systematic comparison of convolutional neural networks (ResNet-18, EfficientNet-B0) and vision transformers (ViT-Tiny, Swin-Tiny) for thermal PV fault detection, using XRAI saliency analysis to assess alignment with thermal physics principles. This represents the first systematic comparison of CNNs and vision transformers for thermal PV fault detection with physics-validated interpretability. Evaluation on 20,000 infrared images spanning normal operation and 11 fault categories shows that Swin Transformer achieves the highest performance (94% binary accuracy; 73% multiclass accuracy) compared to CNN approaches. XRAI analysis reveals that models learn physically meaningful features, such as localized hotspots for cell defects, linear thermal paths for diode failures, and thermal boundaries for vegetation shading, consistent with expected thermal signatures. However, performance varies significantly across fault types: electrical faults achieve strong detection (F1-scores >0.90) while environmental factors like soiling remain challenging (F1-scores 0.20-0.33), indicating limitations imposed by thermal imaging resolution. The thermal physics-guided interpretability approach provides methodology for validating AI decision-making in energy monitoring applications, addressing deployment barriers in renewable energy infrastructure.

The clinical adoption of artificial intelligence (AI) in medical diagnostics is critically hampered by its black-box nature, which prevents clinicians from verifying the rationale behind automated decisions. To overcome this fundamental barrier, we introduce DeepMedix-R1, a foundation model (FM) for chest X-ray (CXR) interpretation that generates not only accurate diagnoses but also a transparent, step-by-step reasoning process grounded in specific visual evidence. Our methodology employs a sequential training strategy, beginning with instruction fine-tuning, followed by a cold-start phase to elicit reasoning capabilities. Critically, we then implement reinforcement learning with grounded rewards to meticulously refine the model, aligning both its diagnostic outputs and its reasoning pathways with clinical plausibility. Quantitative assessments show that DeepMedix-R1 substantially outperforms advanced FMs, achieving improvements in report generation and visual question answering tasks. We also introduce Report Arena, a novel LLM-based benchmark that ranks DeepMedix-R1 first among competing models for output quality. Most significantly, a formal review by clinical experts reveals a profound preference for DeepMedix-R1's generated reasoning over the broadly adopted Qwen2.5-VL-7B model, confirming its superior interpretability and clinical utility.

In this paper, we study the problem of AI explanation of misinformation, where the goal is to identify explanation designs that help improve users' misinformation detection abilities and their overall user experiences. Our work is motivated by the limitations of current Explainable AI (XAI) approaches, which predominantly focus on content explanations that elucidate the linguistic features and sentence structures of the misinformation. To address this limitation, we explore various explanations beyond content explanation, such as "social explanation" that considers the broader social context surrounding misinformation, as well as a "combined explanation" where both the content and social explanations are presented in scenarios that are either aligned or misaligned with each other. To evaluate the comparative effectiveness of these AI explanations, we conduct two online crowdsourcing experiments in the COVID-19 (Study 1 on Prolific) and Politics domains (Study 2 on MTurk). Our results show that AI explanations are generally effective in aiding users to detect misinformation, with effectiveness significantly influenced by the alignment between content and social explanations. We also find that the order in which explanation types are presented - specifically, whether a content or social explanation comes first - can influence detection accuracy, with differences found between the COVID-19 and Political domains. This work contributes towards more effective design of AI explanations, fostering a deeper understanding of how different explanation types and their combinations influence misinformation detection.