Research Papers

Existing paradigms for ensuring AI safety, such as guardrail models and alignment training, often compromise either inference efficiency or development flexibility. We introduce Disentangled Safety Adapters (DSA), a novel framework addressing these challenges by decoupling safety-specific computations from a task-optimized base model. DSA utilizes lightweight adapters that leverage the base model's internal representations, enabling diverse and flexible safety functionalities with minimal impact on inference cost. Empirically, DSA-based safety guardrails substantially outperform comparably sized standalone models across hate speech classification, detecting unsafe model inputs and responses, and hallucination detection with relative improvements of up to 53% in AUC. Furthermore, DSA-based safety alignment allows dynamic, inference-time adjustment of alignment strength and a fine-grained trade-off between instruction following performance and model safety. Importantly, combining the DSA safety guardrail with DSA safety alignment facilitates context-dependent alignment strength, boosting safety on StrongREJECT by 93% while maintaining 98% performance on MTBench - a total reduction in alignment tax of 8 percentage points compared to standard safety alignment fine-tuning. Overall, DSA presents a promising path towards more modular, efficient, and adaptable AI safety and alignment.

Artificial Intelligence (AI) is transforming sectors such as healthcare, finance, and autonomous systems, offering powerful tools for innovation. Yet its rapid integration raises urgent ethical concerns related to data ownership, privacy, and systemic bias. Issues like opaque decision-making, misleading outputs, and unfair treatment in high-stakes domains underscore the need for transparent and accountable AI systems. This article addresses these challenges by proposing a modular ethical assessment framework built on ontological blocks of meaning-discrete, interpretable units that encode ethical principles such as fairness, accountability, and ownership. By integrating these blocks with FAIR (Findable, Accessible, Interoperable, Reusable) principles, the framework supports scalable, transparent, and legally aligned ethical evaluations, including compliance with the EU AI Act. Using a real-world use case in AI-powered investor profiling, the paper demonstrates how the framework enables dynamic, behavior-informed risk classification. The findings suggest that ontological blocks offer a promising path toward explainable and auditable AI ethics, though challenges remain in automation and probabilistic reasoning.

ASyMOB: Algebraic Symbolic Mathematical Operations Benchmark

Michael Shalyt, Rotem Elimelech, Ido Kaminer

Large language models (LLMs) are increasingly applied to symbolic mathematics, yet existing evaluations often conflate pattern memorization with genuine reasoning. To address this gap, we present ASyMOB, a high-resolution dataset of 35,368 validated symbolic math problems spanning integration, limits, differential equations, series, and hypergeometrics. Unlike prior benchmarks, ASyMOB systematically perturbs each seed problem using symbolic, numeric, and equivalence-preserving transformations, enabling a fine-grained assessment of generalization. Our evaluation reveals three key findings: (1) most models' performance collapses under minor perturbations, while top systems exhibit an apparent regime shift in robustness; (2) integrated code tools stabilize performance, particularly for weaker models; and (3) we identify examples where Computer Algebra Systems (CAS) fail while LLMs succeed, as well as problems solved only via a hybrid LLM-CAS approach, highlighting a promising integration frontier. ASyMOB serves as a principled diagnostic tool for measuring and accelerating progress toward building verifiable, trustworthy AI for scientific discovery.

Current jailbreak attacks on large language models (LLMs) predominantly rely on LLMs themselves to generate adversarial prompts, creating a critical efficiency bottleneck: each attack requires substantial computational resources and API queries, limiting scalability and practical deployment. To overcome this limitation, we propose Adversarial Prompt Distillation (APD), a novel framework that transfers jailbreaking capabilities from LLMs to small language models (SLMs) for efficient, low-resource attacks. APD integrates three key components: (1) masked adversarial knowledge pre-training via LoRA fine-tuning, (2) dynamic temperature-controlled knowledge distillation to bridge architectural gaps, and (3) reinforcement learning-based template optimization for adaptive refinement. Extensive experiments across 12 models show that APD achieves state-of-the-art attack success rates (e.g., 96.4% ASR_k on GPT-4) while dramatically improving efficiency - generating prompts 3.7x faster with 11.3x fewer parameters than teacher models. Our work establishes the first practical framework for lightweight jailbreak attacks, exposes new vulnerabilities in LLM defenses, and provides a scalable testbed for advancing AI safety research. Our code is available at: https://github.com/lxgem/Efficient_and_Stealthy_Jailbreak_Attacks_via_Adversarial_Prompt.

Advanced diagnostic instruments are crucial for the accurate detection and treatment of lung diseases, which affect millions of individuals globally. This study examines the effectiveness of deep learning and transfer learning models using a hybrid dataset, created by merging four individual datasets from Bangladesh and global sources. The hybrid dataset significantly enhances model accuracy and generalizability, particularly in detecting COVID-19, pneumonia, lung opacity, and normal lung conditions from chest X-ray images. A range of models, including CNN, VGG16, VGG19, InceptionV3, Xception, ResNet50V2, InceptionResNetV2, MobileNetV2, and DenseNet121, were applied to both individual and hybrid datasets. The results showed superior performance on the hybrid dataset, with VGG16, Xception, ResNet50V2, and DenseNet121 each achieving an accuracy of 99%. This consistent performance across the hybrid dataset highlights the robustness of these models in handling diverse data while maintaining high accuracy. To understand the models implicit behavior, explainable AI techniques were employed to illuminate their black-box nature. Specifically, LIME was used to enhance the interpretability of model predictions, especially in cases of misclassification, contributing to the development of reliable and interpretable AI-driven solutions for medical imaging.

EPIC: Explanation of Pretrained Image Classification Networks via Prototype

Piotr Borycki, Magdalena Trędowicz, Szymon Janusz +4

Explainable AI (XAI) methods generally fall into two categories. Post-hoc approaches generate explanations for pre-trained models and are compatible with various neural network architectures. These methods often use feature importance visualizations, such as saliency maps, to indicate which input regions influenced the model's prediction. Unfortunately, they typically offer a coarse understanding of the model's decision-making process. In contrast, ante-hoc (inherently explainable) methods rely on specially designed model architectures trained from scratch. A notable subclass of these methods provides explanations through prototypes, representative patches extracted from the training data. However, prototype-based approaches have limitations: they require dedicated architectures, involve specialized training procedures, and perform well only on specific datasets. In this work, we propose EPIC (Explanation of Pretrained Image Classification), a novel approach that bridges the gap between these two paradigms. Like post-hoc methods, EPIC operates on pre-trained models without architectural modifications. Simultaneously, it delivers intuitive, prototype-based explanations inspired by ante-hoc techniques. To the best of our knowledge, EPIC is the first post-hoc method capable of fully replicating the core explanatory power of inherently interpretable models. We evaluate EPIC on benchmark datasets commonly used in prototype-based explanations, such as CUB-200-2011 and Stanford Cars, alongside large-scale datasets like ImageNet, typically employed by post-hoc methods. EPIC uses prototypes to explain model decisions, providing a flexible and easy-to-understand tool for creating clear, high-quality explanations.

Psychiatric disorders affect millions globally, yet their diagnosis faces significant challenges in clinical practice due to subjective assessments and accessibility concerns, leading to potential delays in treatment. To help address this issue, we present Heart2Mind, a human-centered contestable psychiatric disorder diagnosis system using wearable electrocardiogram (ECG) monitors. Our approach leverages cardiac biomarkers, particularly heart rate variability (HRV) and R-R intervals (RRI) time series, as objective indicators of autonomic dysfunction in psychiatric conditions. The system comprises three key components: (1) a Cardiac Monitoring Interface (CMI) for real-time data acquisition from Polar H9/H10 devices; (2) a Multi-Scale Temporal-Frequency Transformer (MSTFT) that processes RRI time series through integrated time-frequency domain analysis; (3) a Contestable Diagnosis Interface (CDI) combining Self-Adversarial Explanations (SAEs) with contestable Large Language Models (LLMs). Our MSTFT achieves 91.7% accuracy on the HRV-ACC dataset using leave-one-out cross-validation, outperforming state-of-the-art methods. SAEs successfully detect inconsistencies in model predictions by comparing attention-based and gradient-based explanations, while LLMs enable clinicians to validate correct predictions and contest erroneous ones. This work demonstrates the feasibility of combining wearable technology with Explainable Artificial Intelligence (XAI) and contestable LLMs to create a transparent, contestable system for psychiatric diagnosis that maintains clinical oversight while leveraging advanced AI capabilities. Our implementation is publicly available at: https://github.com/Analytics-Everywhere-Lab/heart2mind.

It is well recognised that ensuring fair AI systems is a complex sociotechnical challenge, which requires careful deliberation and continuous oversight across all stages of a system's lifecycle, from defining requirements to model deployment and deprovisioning. Dynamic argument-based assurance cases, which present structured arguments supported by evidence, have emerged as a systematic approach to evaluating and mitigating safety risks and hazards in AI-enabled system development and have also been extended to deal with broader normative goals such as fairness and explainability. This paper introduces a systems-engineering-driven framework, supported by software tooling, to operationalise a dynamic approach to argument-based assurance in two stages. In the first stage, during the requirements planning phase, a multi-disciplinary and multi-stakeholder team define goals and claims to be established (and evidenced) by conducting a comprehensive fairness governance process. In the second stage, a continuous monitoring interface gathers evidence from existing artefacts (e.g. metrics from automated tests), such as model, data, and use case documentation, to support these arguments dynamically. The framework's effectiveness is demonstrated through an illustrative case study in finance, with a focus on supporting fairness-related arguments.

Safety frameworks have emerged as a best practice for managing risks from frontier artificial intelligence (AI) systems. However, it may be difficult for stakeholders to know if companies are adhering to their frameworks. This paper explores a potential solution: third-party compliance reviews. During a third-party compliance review, an independent external party assesses whether a frontier AI company is complying with its safety framework. First, we discuss the main benefits and challenges of such reviews. On the one hand, they can increase compliance with safety frameworks and provide assurance to internal and external stakeholders. On the other hand, they can create information security risks, impose additional cost burdens, and cause reputational damage, but these challenges can be partially mitigated by drawing on best practices from other industries. Next, we answer practical questions about third-party compliance reviews, namely: (1) Who could conduct the review? (2) What information sources could the reviewer consider? (3) How could compliance with the safety framework be assessed? (4) What information about the review could be disclosed externally? (5) How could the findings guide development and deployment actions? (6) When could the reviews be conducted? For each question, we evaluate a set of plausible options. Finally, we suggest "minimalist", "more ambitious", and "comprehensive" approaches for each question that a frontier AI company could adopt.

Societal cognitive overload, driven by the deluge of information and complexity in the AI age, poses a critical challenge to human well-being and societal resilience. This paper argues that mitigating cognitive overload is not only essential for improving present-day life but also a crucial prerequisite for navigating the potential risks of advanced AI, including existential threats. We examine how AI exacerbates cognitive overload through various mechanisms, including information proliferation, algorithmic manipulation, automation anxieties, deregulation, and the erosion of meaning. The paper reframes the AI safety debate to center on cognitive overload, highlighting its role as a bridge between near-term harms and long-term risks. It concludes by discussing potential institutional adaptations, research directions, and policy considerations that arise from adopting an overload-resilient perspective on human-AI alignment, suggesting pathways for future exploration rather than prescribing definitive solutions.

Responsible Artificial Intelligence (RAI) is a combination of ethics associated with the usage of artificial intelligence aligned with the common and standard frameworks. This survey paper extensively discusses the global and national standards, applications of RAI, current technology and ongoing projects using RAI, and possible challenges in implementing and designing RAI in the industries and projects based on AI. Currently, ethical standards and implementation of RAI are decoupled which caters each industry to follow their own standards to use AI ethically. Many global firms and government organizations are taking necessary initiatives to design a common and standard framework. Social pressure and unethical way of using AI forces the RAI design rather than implementation.

How Large Language Models Are Changing MOOC Essay Answers: A Comparison of Pre- and Post-LLM Responses

Leo Leppänen, Lili Aunimo, Arto Hellas +2

The release of ChatGPT in late 2022 caused a flurry of activity and concern in the academic and educational communities. Some see the tool's ability to generate human-like text that passes at least cursory inspections for factual accuracy ``often enough'' a golden age of information retrieval and computer-assisted learning. Some, on the other hand, worry the tool may lead to unprecedented levels of academic dishonesty and cheating. In this work, we quantify some of the effects of the emergence of Large Language Models (LLMs) on online education by analyzing a multi-year dataset of student essay responses from a free university-level MOOC on AI ethics. Our dataset includes essays submitted both before and after ChatGPT's release. We find that the launch of ChatGPT coincided with significant changes in both the length and style of student essays, mirroring observations in other contexts such as academic publishing. We also observe -- as expected based on related public discourse -- changes in prevalence of key content words related to AI and LLMs, but not necessarily the general themes or topics discussed in the student essays as identified through (dynamic) topic modeling.