British Library

Recent advances in machine learning and AI, including Generative AI and LLMs, are disrupting technological innovation, product development, and society as a whole. AI's contribution to technology can come from multiple approaches that require access to large training data sets and clear performance evaluation criteria, ranging from pattern recognition and classification to generative models. Yet, AI has contributed less to fundamental science in part because large data sets of high-quality data for scientific practice and model discovery are more difficult to access. Generative AI, in general, and Large Language Models in particular, may represent an opportunity to augment and accelerate the scientific discovery of fundamental deep science with quantitative models. Here we explore and investigate aspects of an AI-driven, automated, closed-loop approach to scientific discovery, including self-driven hypothesis generation and open-ended autonomous exploration of the hypothesis space. Integrating AI-driven automation into the practice of science would mitigate current problems, including the replication of findings, systematic production of data, and ultimately democratisation of the scientific process. Realising these possibilities requires a vision for augmented AI coupled with a diversity of AI approaches able to deal with fundamental aspects of causality analysis and model discovery while enabling unbiased search across the space of putative explanations. These advances hold the promise to unleash AI's potential for searching and discovering the fundamental structure of our world beyond what human scientists have been able to achieve. Such a vision would push the boundaries of new fundamental science rather than automatize current workflows and instead open doors for technological innovation to tackle some of the greatest challenges facing humanity today.

IntroductionAdversarial robustness in artificial intelligence is commonly defined in terms of input-level perturbations applied to static models. This study reconceptualises adversarial vulnerability for artificial and agentic AI systems by extending the threat model to autonomy, self-governance, and closed-loop decision-making, where behaviour unfolds dynamically through feedback and control.MethodsWe develop a system-level analytical framework that formalises adversarial risk across perceptual, cognitive, and executive layers. The analysis is grounded in a PRISMA-compliant systematic literature review, bibliometric mapping, and targeted empirical validation. Established adversarial results from vision benchmarks and recent large-language-model red-teaming studies are synthesised to contextualise the framework, rather than to introduce new benchmark performance claims.ResultsThe results demonstrate that no single defence mechanism provides robustness across all layers of agentic AI systems. Adversarial vulnerabilities propagate from perception to policy and actuation, with architectural similarity, domain shift, and feedback dynamics critically shaping transferability and failure modes. These effects have direct implications for safety-critical applications, including autonomous mobility, healthcare imaging, and biometric security.DiscussionBy framing higher-order agentic adversarial threats as hypothesis-driven, system-level risks, this work shifts adversarial AI security from benchmark-centric evaluation to behavioural integrity and lifecycle resilience. The proposed framework defines a coherent research agenda for agentic AI security that integrates control-theoretic reasoning and governance-aware defence design, addressing limitations of classical adversarial machine-learning theory.

Previous scholarship on Southeast Asian Hajj epistolography has treated letters exchanged between Mecca-based Jawi scholars and correspondents in the Malay Archipelago as evidence of an exclusively inter-Muslim network of religious transmission, social relations and economic exchange, occasionally sustained by the trade of Southeast Asian and Arabian goods. A newly identified Malay-language letter in the National Archives of Thailand challenges this model. Found among the papers of Bowon Wichaichan, Siam’s last Deputy King (r. 1868–1885), the 1875 letter from a Malay official in Mecca, Syaikh Salih Daud, acknowledges receipt of valuable trade goods sent via Malay pilgrims, announces the return of Arabian gifts, and requests the donation of white cloth for tents to shelter Muslim pilgrims at Arafat. The correspondence shows that the Siamese Deputy King, a senior member of a Southeast Asian Buddhist dynasty, actively participated in the commercial transactions that accompanied the Hajj, relying on the assistance of Muslim pilgrims and local brokers. It further suggests a legitimate expectation, on the part of the Muslim participants, that this cooperation might translate into royal patronage extended to pilgrims from the Siamese kingdom or its Malay vassal states. This article presents a critical edition, annotated translation, and images of the letter, making this unusual witness to nineteenth-century Siamese Buddhist–Malay Muslim relations available for the first time.

We introduce an extensive qualitative spatial and temporal reasoning (QSTR) benchmark for evaluating large language models (LLMs). We pose questions concerning compositional reasoning (using composition tables, CT), converse relations, and conceptual neighbourhoods (CN) for QSTR calculi, Point Algebra (PA), Allen's Interval Algebra, Interval and Duration (INDU), Region Connection Calculus (RCC-5, RCC-8, and RCC-22), the nine intersection model, cardinal direction calculus, and STAR. The RCC-22 CN is published here for the first time. An extended benchmark systematically varies question presentation including prefix/infix, words/symbols/nonce terms and schematic descriptions for selected calculi. We report results for contemporary frontier models. All models tested perform better than guessing but none can consistently answer all questions correctly. Performance varies sharply by calculus, with PA being the most straightforward, and RCC-22 the most difficult. We release the benchmark, and our results under an open licence to facilitate further assessment of qualitative spatio/temporal reasoning in LLMs.

BACKGROUND:Compositional data comprise the parts of a 'whole' (or 'total'), which sum to that 'whole'. The 'whole' may vary between units of analyses, or it may be fixed (constant). For example, total energy intake (a variable total) is the sum of intake from all foods or macronutrients. Total time in a day (a fixed total) is the sum of time spent engaging in various activities. There exist different approaches to analysing compositional data, such as the isocaloric or isotemporal model, ratio variables, and compositional data analysis (CoDA). Although the performance of the different approaches has been compared previously, this has only been conducted in real data. Since the true relationships are unknown in real data, it is difficult to compare model performance in estimating a known effect. We use data simulations of different parametric relationships, to explore and demonstrate the performance of each approach under various possible conditions. METHODS:We simulated physical activity time-use and dietary data as examples of compositional data with fixed and variable totals, respectively, using different parametric relationships between the compositional components and the outcome (fasting plasma glucose): linear, log2, and isometric log-ratios. We evaluated the performance of a range of generalised linear and additive models as well as CoDA, in estimating a 1-unit and either 10-unit (for physical activity) or 100-unit (for dietary data) reallocations under each parametric scenario. We simulated 10,000 datasets with 1,000 observations in each. RESULTS:The performance of each approach to analysing compositional data depends on how closely its parameterisation matches the true data generating process. Overall, we demonstrated that the consequences of using an incorrect parameterisation (e.g. using CoDA when the true relationship is linear) are more severe for larger reallocations (e.g. 10-min or 100-kcal) than for 1-unit reallocations. The implications of choosing an unsuitable approach may be starker in compositional data with variable totals. For example, while models with ratio variables are mathematically equivalent to linear models in compositional data with fixed totals, their estimates may be radically different for variable totals. CONCLUSIONS:Compositional data with fixed and variable totals behave differently. All existing approaches to analysing such data have utility but need to be carefully selected. Investigators should explore the shape of the relationships between the compositional components and the outcome and chose an approach that matches it best.