Computer Science and Mathematics

Abstract: This paper introduces the concept of Omniscient Mathematics as a formal framework for analyzing the epistemological and computational limits of superintelligent systems operating in spaces of total information. Inspired by Jorge Luis Borges’ Library of Babel, the proposed framework investigates axiomatically environments in which all possible symbolic combinations already exist, including all true statements, false statements, mathematical proofs, scientific theories, and random structures. I demonstrate that the existence of total information does not imply the existence of accessible knowledge. In such environments, exhaustive search becomes computationally and physically infeasible due to combinatorial explosion, semantic entropy, and the impossibility of brute-force verification. Consequently, intelligence emerges not as a mechanism for generating information, but as a necessary process of semantic selection, compression, interpretation, and truth extraction. The paper formally distinguishes structural omniscience from cognitive omniscience, showing that possessing all possible information is fundamentally different from understanding or identifying meaningful knowledge within it. Based on this distinction, we establish the Theorem of the Necessary Emergence of Intelligence, demonstrating that intelligent agents necessarily arise whenever informational spaces become sufficiently large, dense, or combinatorially complete. The proposed theory establishes deep connections between information theory, computability, artificial intelligence, semantic search, and superintelligence. Furthermore, it suggests that future superintelligent systems will not primarily depend on infinite storage or exhaustive computation, but rather on increasingly efficient mechanisms for semantic navigation in spaces of extreme informational complexity. The results provide a new theoretical perspective on the nature of intelligence, knowledge discovery, and the fundamental limits of artificial superintelligence.

Abstract: Rare and safety-critical traffic situations remain challenging for autonomous driving (AD) because they are poorly represented in common training data and may include objects outside standard detector classes. This paper presents a real-time RGB-LiDAR fusion framework for detecting and reacting to rare traffic situations in CARLA. The approach combines YOLOv8-based RGB perception, bird’s-eye-view (BEV) LiDAR clustering, decision-level fusion, an interpretable rule-based safety agent with hysteresis, and an automatic emergency braking (AEB) override. Fused observations are classified as semantic-geometric detections, semantic-only detections, or geometric-only obstacle candidates, where unmatched LiDAR clusters are treated conservatively as candidate-level physical evidence. The framework was evaluated on three CARLA maps and 3CSim-inspired corner-case scenarios, comprising 19253 frames. On a manually annotated subset of 1200 frames, the full pipeline achieved 93.7% precision, 94.7% recall, and a 94.2% F1-score. The CPU implementation processed one frame in 34.7 ms on average, remaining within the 50 ms budget of a 20 Hz simulation tick.

Abstract: Agent identity governance is advancing, though core agent identity and authorization questions remain unresolved: existing frameworks provision, authenticate, authorize, and retire non-human and agentic identities, governing the agent’s identity, credentials, and lifecycle while assuming the composition an agent was approved with remains the composition it runs with. This paper argues that assumption is the open seam. An agent’s effective composition—its tools, data sources, delegated authorities, policies, and child agents—is a runtime supply chain of capability, and that supply chain drifts. We introduce composition drift as the departure of an agent’s effective composition from the terms of its approval, and isolate its most consequential form, compositional drift: the accumulation of individually approved changes into capability that none authorized alone. We formalize this with a two-stage operator: a component-level diff detects that the composition changed (component divergence); a capability-closure stage detects when the change authorized something new (compositional drift)—a qualitative boundary, not a numeric threshold. The contribution is not the observation that approved changes can combine dangerously—long known to authorization security—but a temporal governance model for approved composition drift in agentic systems, linking emergent capability to reauthorization and inventory reconciliation. This drift produces shadow infrastructure: resources provisioned outside any inventory through benign, individually approved pathways. We propose composition attestation, a runtime composition-control layer complementary to identity governance. Paired positive and negative scenarios show the model discriminates, not labels. We bound our claims: the model establishes the phenomenon by construction and claims no deployment efficacy.

Abstract: LLM-based agents increasingly interact with external environments through terminal command execution, yet existing surveys have rarely treated the terminal itself as a primary analytical object. This survey examines terminal agents, namely systems whose task progress depends on iterative command execution, textual feedback, and stateful terminal command interaction, and clarifies their boundaries with adjacent categories such as software-engineering agents, GUI- or browser-based computer-use agents, and CLI-packaged assistants. Through a substrate-centered lens, we systematize the literature around architectures and outer-loop design patterns, competence acquisition through executable environments, command--observation trajectories, and post-training, and evaluation protocols for terminal-mediated capabilities. Across systems, acquisition pipelines, and benchmarks, the synthesis shows that outer-loop design is not an implementation detail but a first-class variable that materially shapes measured performance. The evidence further indicates that terminal competence is multi-dimensional, spanning how agents formulate actions, interpret feedback, manage runtimes, track state and context, verify progress, recover from failures, and control side effects, rather than reducible to a single capability ranking. Current evidence remains concentrated in software engineering, while cross-domain transfer, model-versus-scaffold attribution, reliable recovery in mutable environments, and process-level evaluation remain underdeveloped. The survey provides an evidence-calibrated map of established findings, emerging practices, and unresolved challenges for terminal agents.

Abstract: Intelligence analysts are increasingly able, and required, to consume and interpret outputs generated by artificial intelligence (AI) enabled tools — yet most receive little training in what these outputs actually represent, or how these might be robustly evaluated. This primer addresses that gap. It argues that analysts do not need to know how to develop or operate AI tools in order to use these tools’ outputs critically and judiciously. But they do need sufficient conceptual understanding, and foundational technical knowledge, to evaluate these outputs competently. Three principles provide the framework for this understanding: First, the distinction between AI-facilitated outputs – where automation improves the pace, scale and fidelity of data collection, processing and analytical procedures that analysts could otherwise perform; and AI-generated outputs – where many of the novel outputs generated by the semi-autonomous techniques involved could not have been produced by analysts working independently; Second, the critical difference between interpolative and extrapolative estimation and mechanistic prediction; and Third, the critical dependencies and substantive limitations that govern the reproducibility and practical utility of all AI-facilitated and AI-generated outputs. Together these principles constitute the technical and conceptual foundations of the AI literacy training that all-source intelligence analysts should receive – the case for which is presented in a companion piece to this article.

Abstract: This narrative review examines algorithmic accountability and continuous audit in high-risk public AI systems. Public-sector AI systems are increasingly used in areas such as criminal justice, healthcare, welfare, migration, taxation, education, and public security, where automated or semi-automated outputs may affect rights, access to services, liberty, welfare, and public trust. Rather than presenting original empirical data, this review synthesises interdisciplinary literature, regulatory instruments, standards, and documented public-sector AI cases to identify recurring accountability gaps. The review argues that transparency and compliance-oriented documentation are necessary but insufficient for high-risk public AI systems. Public institutions require audit-ready governance structures capable of linking system design, data provenance, model behaviour, human oversight, monitoring, redress, and institutional responsibility. The article proposes an audit-ready accountability perspective in which algorithmic systems are assessed as sociotechnical infrastructures rather than isolated technical tools. The review concludes that continuous audit, traceability, contestability, and post-deployment monitoring are essential conditions for accountable public-sector AI.

Abstract: Artificial intelligence (AI) is rapidly reshaping how information is collected, processed and analyzed by intelligence agencies. There is therefore a pressing need for foundational AI literacy training to ensure intelligence analysts become confident consumers of the outputs that AI-enabled tools can provide – but not for the specialised computational and statistical training required to become accomplished designers, builders or operators of such tools. AI literacy training should distinguish between: AI-facilitated outputs, where automated collection, processing and analytical workflows can improve the consistency and efficiency of tasks that analysts themselves could perform; and AI-generated outputs, where semi-autonomous computational techniques can prompt unprecedented insights into hitherto, and in some cases previously unknowable, dataset features – insights that nonetheless entail dependencies and limitations requiring careful consideration and evaluation. Drawing on influential policy reports and academic articles exploring the potential utility of AI within intelligence analysis, this paper concludes that the transition to fully AI-enabled analytical capability demands a substantive expansion in AI literacy training; the technical and conceptual foundations of which have been summarized in a companion piece to this article.

Abstract: We present PetroAgents, a multi-agent, multi-modal large-language-model framework for petroleum-engineering reasoning on the Equinor Volve open dataset. The target architecture mirrors an integrated asset team: discipline-specific evidence surfaces, cross-examination, a Council Synthesiser, distributional well-action proposals, and risk review. Read as an agent-design pattern map, the architecture is organised around the seven cognitive functions of a language agent, with the per-discipline evidence lock as its governance layer; four functions are implemented and evaluated in this submission and three are specified as design. The current quantitative evidence is narrower by design. We specify Volve Bench, a six-task Volve benchmark suite spanning DRILL-NPT root-cause attribution, formation top picking, six-month production forecasting, stuck-pipe early warning, multi-modal Discovery-Report QA, and per-wellbore lifecycle forensic analysis over 26 wellbores, 1,759 daily drilling reports, 56 million WITSML rows, 602 LAS files, 5.7 million horizon points, and a decade of production. This submission reports two landed evidence slices: a three-seed DRILL-NPT study on stratified samples drawn from a 1,750-example pool, and a 12-question DISCOVERY-QA smoke test on three rendered pages of the 194-page Hugin Discovery Report. Every reported LLM call goes through a local OpenAI-compatible gateway using locally-hosted open weights (GPT-OSS-120B, Qwen3.6-35B, Gemma-4-31B, MiniMax-M2.7, Qwen3-VL-235B-FP8); no paid frontier API is invoked. On DRILL-NPT, the four-family vote attains macro-F1 0.464 ± 0.012 on the broad all-wellbore sample and 0.442 ± 0.019 on the WITSML-applicable subset. The Drilling+HSE+Council path lifts WITSML-applicable macro-F1 by +0.048 over the single-LLM baseline, but the paired test is not significant at three seeds (p = 0.22), so we report it as directional evidence rather than a settled win. A same-subset evidence-redaction ablation shows that exposing state_detail and proprietary_code lifts B1 from 0.355 to 0.431 macro-F1, quantifying how much of DRILL-NPT is label-code leakage rather than prose reasoning. On DISCOVERY-QA, Qwen3-VL reading rendered page images reaches a 0.958 keyword-hit score versus 0.792 for GPT-OSS-120B reading pdftotext, a bounded +16.7 percentage-point lift concentrated on figure annotations and OCR-damaged numerics.

Abstract: There is a worldwide danger of fire incidents that can lead to significant destruction of lives and property, especially in urban and smart cities. Existing fire and smoke detection systems are often inadequate for detecting the location of a fire, assessing the speed of its spread, and providing real-time alerts that can be acted upon quickly. This study proposes a method termed SmartFire Vision, which uses a hybrid deep-learning framework consisting of an Efficient Vision Transformer (E-ViT) and a Detection Transformer (DETR) for real-time fire and smoke detection from video sequences. A major contribution of this study is the integration of a new Removing Inefficient Attention Heads (RIAH) pruning strategy to reduce the computational overhead and maintain a global context in the ViT encoder. The E-ViT and DETR feature representations were fused and passed to a fully connected classification head enhanced with a probabilistic thresholding function and an integrated alarm system. The proposed model was trained and evaluated using the FURG fire benchmark dataset, which comprises 28,022 annotated frames. The proposed model achieved an overall accuracy of 91.37%, precision of 88.55%, recall of 85.27%, and F1-score of 86.64%, surpassing the current state-of-the-art methods. The SmartFire Vision framework provides a highly capable and computationally efficient means of fire detection and is particularly beneficial for CCTV-based smart city surveillance, with a high likelihood of deployment in edge devices in the future.

Abstract: State-level health resource systems require timely and robust prediction models; nonetheless, they face the persistent issue of declining model effectiveness due to continuously changing data distributions (data drift). This issue is particularly pronounced in markedly imbalanced classification tasks, such as predicting rare yet critical events like Sickle Cell Crisis. This research introduces an Evolving AI-Driven Ensemble Learning Framework that integrates advanced ensemble methods (stacking XGBoost, Deep Neural Network, and Random Forest with a meta-learner) alongside novelty detection employing the F1-Score. Sophisticated feature engineering techniques, including automated feature selection through evolutionary algorithms and meta-learning (MAML), are employed to handle complex, high-dimensional health data. We simulated real-time data drift and performed empirical assessments on a reactive re-training method enhanced by ensemble stacking. The technique successfully validated its novelty detection, as the F1-Score consistently and significantly remained below the adaptation threshold in response to drift. The enhanced ensemble and feature engineering mitigated the shortcomings of basic adaption methods, resulting in a 10% gain in F1-Score and a 15% improvement in Precision compared to the original model after five re-training iterations. This outcome demonstrates that optimised ensemble learning and automated feature engineering improve model stability and maintain accuracy for minority classes in complex data environments. A resolved runtime issue in the SHAP explainability layer improves model transparency in pipeline development. This study substantiates the necessity for intelligent control systems, specifically Meta-Learning for precise feature updates and Reinforcement Learning (RL) for the dynamic development of optimal adaptation policies, thus ensuring the framework's effectiveness as a dependable and truly autonomous system in public health.

Abstract: Purpose: To develop and evaluate a deep learning-based quality control system for Lumbar Spinal Digital Radiographs (LSDR), designed to automate and improve their evaluation and reduce reliance on manual reviews. Methods: This retrospective study utilized a deep learning workflow comprising image segmentation, feature extraction, and a classification model. The dataset, including anteroposterior (AP) and lateral (LAT) X-ray images, was expanded through data augmentation techniques. Four U-Net-based models were assessed: standard U-Net, Swin-UNet, Attention U-Net, and Attention U-Net with weighted attention, with the latter selected for its superior performance. Extracted features, such as brightness, contrast, and anatomical positioning, were used in an XGBoost classifier, which was evaluated using mean intersection over union (mIoU), accuracy, sensitivity, specificity, and AUC. Results: The Attention U-Net with weighted attention outperformed the other models, achieving high mIoU scores in both AP and LAT views. The XGBoost classifier achieved the best performance in classifying images as “qualified” or “unqualified,” with an AUC of approximately 0.9, high accuracy, and balanced sensitivity and specificity. This approach effectively addressed class imbalances and improved model accuracy compared to traditional machine learning models such as MLP and SVM. Conclusion: The developed automated quality control system demonstrated potential for enhancing image quality, enhancing diagnostic reliability, and optimizing clinical workflow efficiency.

Abstract: Feature preprocessing is standard practice in machine learning, but how much it actually matters depends on the model and the data. We investigate this question by evaluating five preprocessing strategies (no preprocessing, min-max scaling, standardisation, PCA, and whitening) across five classical classifiers on a phishing URL dataset with over 235,000 samples and 50 numerical features. Using stratified five-fold cross-validation and paired t-tests, we find that most models achieve near-perfect performance regardless of preprocessing. The RBF-SVM tells a different story: without scaling its ROC-AUC sits at 0.997, and a controlled scale-distortion experiment pushes it down to 0.532, barely above random chance (p < 0.001). Any scaling method fully restores it. We also find that k-NN benefits from standardisation but not from min-max scaling, that Na¨ıve Bayes is harmed by PCA, and that the ranking of important features changes entirely depending on whether the data is scaled.

Abstract: Detectors built on transformer language models report near-perfect accuracy on standard fake-news benchmarks, which suggests the task is almost solved. We argue that much of this accuracy reflects a confound between writing register and veracity: in common benchmarks the real class is human-written while the fake class is machine-generated or machine-rewritten, so a detector can separate the classes by recognizing AI writing style rather than by judging truth. To test this, we designed a two-regime evaluation. Phase 1 is the standard setup, comparing untouched human-real articles against laundered fake articles. Phase 2 is register-controlled: the real class is passed through the same cross-LLM laundering chains as the fake class, so both classes share one machine register and only veracity separates them. We train seven detectors on three datasets and evaluate each frozen detector under both regimes. Under Phase 1 detectors appear robust; under Phase 2, detection on WELFake collapses from about 99% to about 62% AUROC and the largest models approach chance. The effect is benchmark dependent, large on WELFake, mild on IFND and near zero on GossipCop, and it is confirmed by bootstrap testing with false discovery rate control. We recommend register-controlled evaluation as standard reporting practice.

Abstract: This study investigates a deterministic-first/learned-second AI/ML framework that is deployed in regulated retail brokerage accounts. A two-stage calibration system is employed that combines historical back-testing with live recalibration via continuous position snapshots. The framework includes a unique explainability layer that employs the global sensitivity analysis method, SimDec, to identify the most influential components. The use of SimDec renders all AI-created solutions free of hallucination and fully explainable. An intraday options trading financial framework is illustrated through a live training-investment cycle on long-call positions using “real money”. The major contribution of this research is the overall architectural and methodological framework. The filter-before-you-solve approach enables contributions to be evaluated independently of specific implementations. Beyond financial applications, it is described how the complete architectural pattern can actually generalize to many AI/ML deployment contexts that require auditable deterministic gating prior to learned inference.

Abstract: LLM-based agents mark a shift from passive question answering to active task completion: they perceive environments, invoke tools, maintain state, and act over extended horizons. As agent systems have evolved from prompt engineering to workflows and context engineering, harness engineering, and agent-native training with co-evolution, a central question has become increasingly important: where does the bottleneck in agent performance reside—in the foundation model, in the execution harness, or in the coupling between them? This survey examines LLM-based agents through a model harness lens. We first clarify the functional definition of agents and the implementation view of an LLM-based agent as a foundation model coupled with an execution harness. Wethen analyze the limits of model-centric scaling, trace four paradigms of agent engineering, and decompose the execution harness into six coupled runtime responsibilities: observation, context, control, action, state, and verification/governance. Using this decomposition, we map task properties and domain pressures to harness configurations, review benchmark and evaluation practices, and synthesize model–harness evidence on how runtime design affects long-horizon task completion, efficiency, and reliability. Finally, we identify open challenges in value-aware evaluation, safety, harness generalization, and model–harness co-evolution. Rather than treating agents as models with auxiliary tools, this survey argues that agent quality—including success, efficiency, safety, and generalization—emerges from the interaction between model capability, runtime infrastructure, task structure, and evaluation design. A collection of papers discussed in this survey is provided in https://github.com/ggjy/Awesome-Agent-Engineering.

Abstract: AI classification systems are increasingly used in critical public services to prioritise patients, screen welfare claims, assess risk, route migration applications, detect fraud, allocate educational support, and identify security alerts. Although these systems may improve triage, consistency, and administrative capacity, they also redistribute suspicion, delay access to essential services, affect liberty and welfare, and shift evidential burdens onto citizens. This review examines ethical and accountability challenges in public-sector AI classification through the linked requirements of fairness, explainability, auditability, cybersecurity, data governance, procurement control, human review, and contestability. It contributes a Public Classification Accountability Chain that connects seven auditable elements: public-purpose justification, data and label legitimacy, model and threshold choice, human oversight, audit evidence, citizen contestability, and post-deployment authority. The article argues that high-stakes public classifiers should be governed as sociotechnical accountability systems rather than as isolated predictive tools. Responsible deployment requires public institutions to justify the classification task, document data and thresholds, monitor subgroup harms, reconstruct decisions, secure systems, provide meaningful review and appeal, and suspend or redesign classifiers when their effects cannot be made proportionate, lawful, and accountable.

Abstract: The growing affordability, autonomy, and swarming of small unmanned aerial vehicles (UAVs) turn low-altitude defense from single-shot interception into a multi-node cooperative decision problem, in which the loss of sensing, coordination, or engagement nodes breaks the closed loops linking them. This study formulates their recovery as the dynamic reconfiguration of cooperative counter-UAV task chains. Given a pre-disturbance plan and a set of failed defending nodes, reconfiguration is modeled as a constrained bi-objective optimization balancing recovered engagement effectiveness against the change to the baseline plan, and is solved by Multi-Agent Heuristic Evolution (MAHE), an automated heuristic design framework whose evolution, coordinator, repair, and reflection agents—driven by a large language model—evolve scoring heuristics for a fixed reconfiguration solver. Across instances of varying scale and under light-to-heavy node loss, MAHE outperforms both a single-agent heuristic-design counterpart and a range of hand-crafted solvers: the latter lose most of their solution quality as the problem grows, whereas MAHE preserves it and sustains high recovery at a nearly constant reconfiguration cost; an ablation confirms that its agents contribute complementary gains. These results indicate that automatically generated, reconfiguration-specific heuristics provide a scalable route to dynamic, heterogeneous, and constraint-intensive counter-UAV task planning.

Abstract: Artificial Intelligence (AI) based support systems are transforming the educational landscape by enhancing teaching efficiency, personalized learning, and accessibility. Despite rapid technological progress, educational institutions face persistent challenges such as unequal access to quality learning resources, limited teacher support, and the need for individualized student engagement. These issues hinder effective learning outcomes and inclusivity in modern classrooms. This paper explores the design and implementation of sustainable and AI-based educational support systems that address these challenges through intelligent tutoring, adaptive learning analytics, and automated feedback mechanisms. By integrating natural language processing, machine learning, and predictive modelling, the proposed framework provides real-time assistance to educators and learners, fostering data-driven decision-making and inclusive pedagogy. Qualitative research demonstrates that AI-driven systems can significantly improve academic performance, teacher productivity, and learner motivation, offering a scalable and equitable solution for the future of education in both traditional and digital environments.

Abstract: Modern higher education institutions increasingly rely on interconnected digital platforms, including Learning Management Systems (LMS), digital libraries, academic portals, and human resource information systems, creating significant challenges in user authentication management. This study aimed to design and evaluate an Intelligent Single Sign-On (SSO) Framework based on OAuth 2.0/OpenID Connect (OIDC) within a simulated portal ecosystem reflecting Universitas Negeri Medan (UNIMED). A synthetic simulation was conducted using 180 user scenarios representing students (71.1%), lecturers (17.2%), and administrative staff (11.7%) from nine academic units. The proposed framework integrated OAuth 2.0/OIDC protocols with risk-based adaptive authentication and Role-Based Access Control (RBAC). The simulation results indicated a potential 72.3% reduction in login time, from 48.95 seconds to 13.54 seconds, an increase in authentication success rate from 83.75% to 94.14%, and a decrease in login failures from 3.38 to 0.85 per month. The System Usability Scale (SUS) score improved from 66.21, categorized as marginal, to 85.99, categorized as excellent, with 61.7% of user scenarios reaching the highest category under the post-SSO condition. These findings suggest that the proposed OAuth 2.0/OIDC-based framework provides a strong foundation for improving secure, efficient, and user-centered authentication in higher education institutions.

Abstract: (1) Background: Asthma is a chronic respiratory disease shaped by environmental, meteorological, and behavioral factors. Although the literature has advanced in predicting respiratory outcomes and in deploying digital technologies for therapeutic support, few approaches integrate population surveillance and individual monitoring within the same analytical framework. (2) Methods: This work developed an integrated machine learning framework composed of two complementary studies. Study 1 modeled the daily count of hospital admissions for all respiratory diseases (chapter X of the ICD-10, which includes asthma, COPD, and respiratory tract infections), denoted HOSPCIDX, in the municipality of São Paulo (Brazil). It drew on six years of data (2017 to 2022) from SIH/SUS via PCDaS/Fiocruz, CETESB, and INMET, and applied a hybrid architecture combining ElasticNet, residual CatBoost, direct CatBoost, and adaptive blending, validated through walk-forward over 30 bimonthly folds across the 2018 to 2022 period (2017 was reserved for lag construction). Study 2 focused specifically on pediatric asthma and analyzed 913 qualified records from the Respire Bem system, collected from patients aged 6 to 16 years, using XGBoost and Random Forest models. The clinical outcomes (Asthma Control Test and salivary cortisol) were generated by evidence-based synthetic simulation. (3) Results: In Study 1, the hybrid model achieved a mean MAE of 18.22, a mean RMSE of 23.99, a mean R² of 0.675, and a mean skill gain of 41.5% over the seasonal baseline. SHAP analysis identified mean temperature, PM2.5, NO2, and CO as the main predictive drivers of respiratory hospitalizations. In Study 2, XGBoost reached an R² of 0.80 for the simulated Asthma Control Test and 0.78 for simulated salivary cortisol, with self-reported sentiment emerging as the leading digital biomarker. (4) Conclusions: The proposed framework demonstrates the feasibility of a dual analytical architecture for asthma management, combining environmental prediction at the population level with digital monitoring at the individual level. Study 1 provides robust predictive validation with real data, while Study 2 represents an exploratory stage based on real behavioral data and simulated clinical outcomes, which calls for prospective validation with direct clinical and biological measurements.