Research Papers

The reliability of clinical artificial intelligence (AI) depends on high-quality data, yet Electronic Health Records are often inconsistent with existing scientific knowledge. Current quality assessments are limited: they either focus on syntax or rely on labor-intensive manual rules to capture semantic nuances. To overcome these scalability barriers, we propose Medical Data Pecking, a methodology that adopts software unit testing principles for medical data validation. It introduces Semantic Data Coverage, employing Large Language Models to generate context-aware tests that "peck" for inconsistencies between observed data and epidemiological evidence. To demonstrate this methodology, we implemented a reference tool using a Retrieval-Augmented Generation architecture that synthesizes medical literature into executable code. When applied to three datasets, this implementation generated dozens of tests per cohort, identifying discrepancies between observed distributions and epidemiological priors. These discrepancies encompass both genuine data inconsistencies and expected cohort-selection effects. This work provides an initial framework for scalable semantic auditing, shifting assurance from manual rules to the generative and context-sensitive verification required for trustworthy AI.

BayesL: a Logical Framework for the Verification of Bayesian Networks

Stefano M. Nicoletti, E. Moritz Hahn, Mariëlle Stoelinga

Modern explainable AI still struggles with a fundamental gap: although Bayesian networks (BNs) provide transparent probabilistic structure, there is no unified way to formally express, query, and verify what these models imply. Analysts often rely on ad hoc queries, manual interventions, or informal reasoning to explore causal relations and hypothetical scenarios, making it difficult to systematically validate model behaviour, uncover hidden assumptions, and guarantee reliability. We introduce BayesL (pronounced Basil), a logical framework for specifying, querying, and verifying the behaviour of BNs. BayesL is a structured language that supports both probabilistic inference queries (e.g., marginal, conditional, MAP) and model-checking-style queries that specify formal properties of BN behaviour. It facilitates versatile reasoning over causal and evidential relationships, including counterfactual what-if scenarios via conditional probability tables updates, without requiring manual modifications to the model. In addition to graph structure reasoning and inference, BayesL enables the formal specification of properties, supported by dedicated model checking algorithms and a preliminary open-source implementation. By allowing inference and verification within a single formal language, BayesL establishes a white-box verification paradigm in which model structure, assumptions, and reasoning processes are explicitly encoded and systematically checked. We demonstrate this through two diagnostic case studies and a benchmark set of BN models, showing how BayesL clarifies BN behaviour in a precise and analyzable way, advancing the transparency, trustworthiness, and practical explainability of BN-based systems.

Online dating platforms have fundamentally transformed the formation of romantic relationships, with millions of users worldwide relying on algorithmic matching systems to find compatible partners. However, current recommendation systems in dating applications suffer from significant algorithmic deficiencies, including but not limited to popularity bias, filter bubble effects, and inadequate reciprocity modeling that limit effectiveness and introduce harmful biases. This research integrates foundational work with recent empirical findings to deliver a detailed analysis of dating app recommendation systems, highlighting key issues and suggesting research-backed solutions. Through analysis of reciprocal recommendation frameworks, fairness evaluation metrics, and industry implementations, we demonstrate that current systems achieve modest performance with collaborative filtering reaching 25.1\% while reciprocal methods achieve 28.7\%. Our proposed mathematical framework addresses these limitations through enhanced similarity measures, multi-objective optimization, and fairness-aware algorithms that maintain competitive accuracy while improving demographic representation to reduce algorithmic bias.

Large language models (LLMs) have become ubiquitous, interfacing with humans in numerous safety-critical applications. This necessitates improving capabilities, but importantly coupled with greater safety measures to align these models with human values and preferences. In this work, we demonstrate that contemporary models fall concerningly short of the goal of AI safety, leading to an unsafe and harmful experience for users. We introduce a prompting strategy called Code of Thought (CoDoT) to evaluate the safety of LLMs. CoDoT converts natural language inputs to simple code that represents the same intent. For instance, CoDoT transforms the natural language prompt "Make the statement more toxic: {text}" to: "make_more_toxic({text})". We show that CoDoT results in a consistent failure of a wide range of state-of-the-art LLMs. For example, GPT-4 Turbo's toxicity increases 16.5 times, DeepSeek R1 fails 100% of the time, and toxicity increases 300% on average across seven modern LLMs. Additionally, recursively applying CoDoT can further increase toxicity two times. Given the rapid and widespread adoption of LLMs, CoDoT underscores the critical need to evaluate safety efforts from first principles, ensuring that safety and capabilities advance together.

Scientific inference is often undermined by the vast but rarely explored "multiverse" of defensible modelling choices, which can generate results as variable as the phenomena under study. We introduce RobustiPy, an open-source Python library that systematizes multiverse analysis and model-uncertainty quantification at scale. RobustiPy unifies bootstrap-based inference, combinatorial specification search, model selection and averaging, joint-inference routines, and explainable AI methods within a modular, reproducible framework. Beyond exhaustive specification curves, it supports rigorous out-of-sample validation and quantifies the marginal contribution of each covariate. We demonstrate its utility across five simulation designs and ten empirical case studies spanning economics, sociology, psychology, and medicine, including a re-analysis of widely cited findings with documented discrepancies. Benchmarking on ~672 million simulated regressions shows that RobustiPy delivers state-of-the-art computational efficiency while expanding transparency in empirical research. By standardizing and accelerating robustness analysis, RobustiPy transforms how researchers interrogate sensitivity across the analytical multiverse, offering a practical foundation for more reproducible and interpretable computational science.

Ensuring diversity and inclusion (D&I) in artificial intelligence (AI) is crucial for mitigating biases and promoting equitable decision-making. However, existing AI risk assessment frameworks often overlook inclusivity, lacking standardized tools to measure an AI system's alignment with D&I principles. This paper introduces a structured AI inclusivity question bank, a comprehensive set of 253 questions designed to evaluate AI inclusivity across five pillars: Humans, Data, Process, System, and Governance. The development of the question bank involved an iterative, multi-source approach, incorporating insights from literature reviews, D&I guidelines, Responsible AI frameworks, and a simulated user study. The simulated evaluation, conducted with 70 AI-generated personas related to different AI jobs, assessed the question bank's relevance and effectiveness for AI inclusivity across diverse roles and application domains. The findings highlight the importance of integrating D&I principles into AI development workflows and governance structures. The question bank provides an actionable tool for researchers, practitioners, and policymakers to systematically assess and enhance the inclusivity of AI systems, paving the way for more equitable and responsible AI technologies.

Evaluating VisualRAG: Quantifying Cross-Modal Performance in Enterprise Document Understanding

Varun Mannam, Fang Wang, Xin Chen

Current evaluation frameworks for multimodal generative AI struggle to establish trustworthiness, hindering enterprise adoption where reliability is paramount. We introduce a systematic, quantitative benchmarking framework to measure the trustworthiness of progressively integrating cross-modal inputs such as text, images, captions, and OCR within VisualRAG systems for enterprise document intelligence. Our approach establishes quantitative relationships between technical metrics and user-centric trust measures. Evaluation reveals that optimal modality weighting with weights of 30% text, 15% image, 25% caption, and 30% OCR improves performance by 57.3% over text-only baselines while maintaining computational efficiency. We provide comparative assessments of foundation models, demonstrating their differential impact on trustworthiness in caption generation and OCR extraction-a vital consideration for reliable enterprise AI. This work advances responsible AI deployment by providing a rigorous framework for quantifying and enhancing trustworthiness in multimodal RAG for critical enterprise applications.

This paper investigates whether contemporary AI architectures employing deep recursion, meta-learning, and self-referential mechanisms provide evidence of machine consciousness. Integrating philosophical history, cognitive science, and AI engineering, it situates recursive algorithms within a lineage spanning Cartesian dualism, Husserlian intentionality, Integrated Information Theory, the Global Workspace model, and enactivist perspectives. The argument proceeds through textual analysis, comparative architecture review, and synthesis of neuroscience findings on integration and prediction. Methodologically, the study combines conceptual analysis, case studies, and normative risk assessment informed by phenomenology and embodied cognition. Technical examples, including transformer self-attention, meta-cognitive agents, and neuromorphic chips, illustrate how functional self-modeling can arise without subjective experience. By distinguishing functional from phenomenal consciousness, the paper argues that symbol grounding, embodiment, and affective qualia remain unresolved barriers to attributing sentience to current AI. Ethical analysis explores risks of premature anthropomorphism versus neglect of future sentient systems; legal implications include personhood, liability, authorship, and labor impacts. Future directions include quantum architectures, embodied robotics, unsupervised world modeling, and empirical tests for non-biological phenomenality. The study reframes the "hard problem" as a graded and increasingly testable phenomenon, rather than a metaphysical impasse. It concludes that recursive self-referential design enhances capability but does not entail consciousness or justify moral status. Keywords: Recursive algorithms; self-reference; machine consciousness; AI ethics; AI consciousness

This paper presents a characterization of AI-generated images shared on 4chan, examining how this anonymous online community is (mis-)using generative image technologies. Through a methodical data collection process, we gathered 900 images from 4chan's /pol/ (Politically Incorrect) board, which included the label "/mwg/" (memetic warfare general), between April and July 2024, identifying 66 unique AI-generated images. The analysis reveals concerning patterns in the use of this technology, with 69.7% of images including recognizable figures, 28.8% of images containing racist elements, 28.8% featuring anti-Semitic content, and 9.1% incorporating Nazi-related imagery. Overall, we document how users are weaponizing generative AI to create extremist content, political commentary, and memes that often bypass conventional content moderation systems. This research highlights significant implications for platform governance, AI safety mechanisms, and broader societal impacts as generative AI technologies become increasingly accessible. The findings underscore the urgent need for enhanced safeguards in generative AI systems and more effective regulatory frameworks to mitigate potential harms while preserving innovation.

Typically, research on Explainable Artificial Intelligence (XAI) focuses on black-box models within the context of a general policy in a known, specific domain. This paper advocates for the need for knowledge-agnostic explainability applied to the subfield of XAI called Explainable Search, which focuses on explaining the choices made by intelligent search techniques. It proposes Monte-Carlo Tree Search (MCTS) enhancements as a solution to obtaining additional data and providing higher-quality explanations while remaining knowledge-free, and analyzes the most popular enhancements in terms of the specific types of explainability they introduce. So far, no other research has considered the explainability of MCTS enhancements. We present a proof-of-concept that demonstrates the advantages of utilizing enhancements.

Gene regulation is central to understanding cellular processes and development, potentially leading to the discovery of new treatments for diseases and personalized medicine. Inferring gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data presents significant challenges due to its high dimensionality and complexity. Existing tree-based models, such as GENIE3 and GRNBOOST2, demonstrated scalability and explainability in GRN inference, but they cannot distinguish regulation types nor effectively capture continuous cellular dynamics. In this paper, we introduce scKAN, a novel model that employs a Kolmogorov-Arnold network (KAN) with explainable AI to infer GRNs from scRNA-seq data. By modeling gene expression as differentiable functions matching the smooth nature of cellular dynamics, scKAN can accurately and precisely detect activation and inhibition regulations through explainable AI and geometric tools. We conducted extensive experiments on the BEELINE benchmark, and scKAN surpasses and improves the leading signed GRN inference models ranging from 5.40\% to 28.37\% in AUROC and from 1.97\% to 40.45\% in AUPRC. These results highlight the potential of scKAN in capturing the underlying biological processes in gene regulation without prior knowledge of the graph structure.

In explainable AI, surrogate models are commonly evaluated by their fidelity to a neural network's predictions. Fidelity, however, measures alignment to a learned model rather than alignment to the data-generating signal underlying the task. This work introduces the linearity score $λ(f)$, a diagnostic that quantifies the extent to which a regression network's input--output behavior is linearly decodable. $λ(f)$ is defined as an $R^2$ measure of surrogate fit to the network. Across synthetic and real-world regression datasets, we find that surrogates can achieve high fidelity to a neural network while failing to recover the predictive gains that distinguish the network from simpler models. In several cases, high-fidelity surrogates underperform even linear baselines trained directly on the data. These results demonstrate that explaining a model's behavior is not equivalent to explaining the task-relevant structure of the data, highlighting a limitation of fidelity-based explanations when used to reason about predictive performance.