Computer Science and Mathematics

Abstract: Construction companies, petrochemical companies, and airports are classified as large-scale organizational and technical systems. In organizational and technical system is implemented using hierarchical distributed control systems. To achieve goals, numerous parallel technological operations and business processes are executed, requiring synchronization and consuming significant energy resources. To optimize energy resources at all hierarchical levels, an up-to-date picture of the system state is necessary. This paper proposes constructing a snapshot of the system's state, which allows for an assessment of system performance within selected criteria and the selection of appropriate solutions within the entire control system. Procedures for constructing a snapshot are discussed. An example of optimizing the energy consumption of a control system based on snapshot analysis is provided.

Abstract: Foundation-model agents now use reusable skills for tool use, long-horizon planning, and adaptation across related tasks. The term, however, is used loosely. It may describe a prompt package, an executable workflow, a learned routine, or an artifact distributed through a repository. That looseness makes it hard to compare methods, measure progress, or discuss security and governance with precision.We study agent skills as reusable and adaptive units of competence between model capability and situated task execution. The survey separates skills from nearby constructs such as prompts, tools, memory, and policies, then organizes the literature around representation, lifecycle and orchestration, evaluation, security and governance, and application domains. Across these areas, skill quality is only one part of the story. Useful skills also depend on abstraction choices, retrieval and composition mechanisms, ecosystem structure, and infrastructure security. We treat agent skills as a research object in their own right and identify open problems in automatic induction, cross-environment transfer, longitudinal evaluation, and trustworthy sharing in open agent ecosystems. A public paper list is available at https://github.com/JinhaoShen/awesome-agent-skill-papers.

Abstract: Binary sequences (binary codes), where the elements are −1 or +1, are useful in many fields, including communications, radar, sonar, mathematics, physics, and cryptography. This paper considers binary sequences with low aperiodic autocorrelations and focuses on the small peak sidelobe levels alongside the merit factor. Two families of binary sequences are considered, namely Rudin-Shapiro and Legendre sequences. For both families, we applied a heuristic algorithm to minimize the peak sidelobe levels for sequences of lengths up to 2^16 and 220−1, respectively. The main contribution of the article is two conjectures associated with Legendre sequences: (1) The obtained binary sequences with the best-known peak sidelobe levels have merit factor ≈5.0, (2) The number of elements that differ between the resulting binary sequences and the initial Legendre sequences follows a linear dependence on the sequence length (n), namely ≈0.01n. The Rudin-Shapiro sequences do not exhibit these properties, as worse peak sidelobe level and merit factor values were obtained. The number of elements that differ between the resulting binary sequences and the initial Rudin-Shapiro sequences is also much higher compared to that of the Legendre sequences.

Abstract: Autoregressive large language models (AR-LLMs) have achieved remarkable success, but their inherently sequential decoding process remains a fundamental bottleneck for efficient inference. Diffusion large language models (DLLMs), with bidirectional modeling and parallel token generation, offer a promising alternative to break this token-by-token limitation. Yet despite rapid progress, the practical inference efficiency of current DLLMs remains unclear. From a verification perspective, this survey establishes a systematic taxonomy of existing acceleration methods, benchmarks representative techniques under a unified experimental setting, and further evaluates strong strategy combinations to quantify the gap between mainstream DLLM inference methods and state-of-the-art AR baselines. Specially, the overall analysis highlights that the parallel decoding efficiency of DLLMs still remains a significant lag compared to the decoding efficiency of AR-LLMs under inference acceleration. We provide an in-depth experimental analysis about the underlying trade-offs among generation quality, latency, and system compatibility, and build up a standard evaluation bench open to the community. Remaining bottlenecks are also summarized, together with future directions for more practical and competitive DLLM inference. Code is available at \url{https://github.com/haoyun-jiang/DLLM-AccelEval}.

Abstract: Multi-server private information retrieval (PIR) based on function secret sharing (FSS) has emerged as a prominent paradigm for achieving sublinear communication. However, standard FSS constructions strictly require full server participation, making them highly vulnerable to single-node fail-stop faults. Existing fault-tolerant schemes mitigate this but inevitably inflate the downlink response overhead to scale with the database size N (e.g., \( O(\sqrt{N}) \)). To overcome this limitation, we propose a (t,p)-fault-tolerant PIR (FT-PIR) protocol grounded in a newly designed generalized (t,p)-fault-tolerant distributed point function (FT-DPF). By introducing a hierarchical recursive patching mechanism, our scheme transforms rigid all-party evaluations into flexible t-out-of-p reconstructions. This architecture completely decouples the response communication from N and ensures efficient client-side reconstruction via lightweight XOR aggregations, fundamentally bypassing heavy algebraic interpolations. Formal analysis proves that our strictly stateless protocol guarantees (t-1)-computational privacy under the semi-honest model. Asymptotic evaluations demonstrate that the proposed FT-PIR achieves an optimal downlink complexity bounded to O(\( poly(t,p) \cdot \log p \)), significantly outperforming existing robust baselines for large-scale datasets.

Abstract: Digital Twin (DT) systems are revolutionizing modern industry by enabling real-time monitoring, simulation, and predictive control of physical assets. However, their widespread adoption in critical domains is contingent upon the trust and security they inspire. This paper presents a comprehensive survey of trust and security in DT systems, synthesizing recent advancements to bridge interdisciplinary gaps. We propose a novel taxonomy that categorizes trust into behavioral and non-behavioral dimensions and aligns these with the architectural layers of a DT. The survey meticulously analyzes the evolving threat landscape, detailing DT-specific vulnerabilities and their implications across diverse application domains. Furthermore, we explore current defense mechanisms, architectural models for secure data distribution, and privacy-preserving techniques such as federated learning and differential privacy. The paper also investigates trust-building strategies, including certification, explainable AI, and stakeholder-centric design. Finally, we identify critical open challenges and outline promising future research directions, including the need for unified trust metrics, lightweight security for edge DTs, and resilient, adaptive autonomy. This survey serves as a foundational reference for researchers and practitioners aiming to develop intelligent, connected, and inherently trustworthy digital twin ecosystems.

Abstract: The audit of Information Security Management Systems (ISMS) under ISO/IEC 27001:2022 has traditionally relied on human auditors whose competence, experience, and judgment shape audit outcomes. While effective, this human-centric approach suffers from inter-auditor variability, high cost, scheduling constraints, and limited scalability — challenges magnified by the post-pandemic shift toward remote audits and the growing volume of organisations seeking certification. Recent advances in Natural Language Processing (NLP), Computer Vision (CV), and Large Language Models (LLMs) suggest that significant portions of the audit workflow could be augmented by machine learning. However, prior research has examined these technologies in isolation; no integrated conceptual framework yet exists that unifies document review, field observation, and interviewing under a single multi-modal pipeline tailored to ISO/IEC 27001 audits and explicitly grounded in the audit methodology of ISO 19011:2018. This paper proposes such a framework — the Multi-Modal ML-Augmented ISO 27001 Audit Framework (M³A-Framework). We synthesise insights from ISO 19011:2018 audit guidelines, recent advances in AI-driven assurance, and the design science research paradigm to develop a five-stage conceptual model that augments the seven-step evidence-collection process specified in ISO 19011 Clause 6.4.7 and that extends the audit-methods matrix of ISO 19011 Annex A (Table A.1). The framework comprises: (1) audit planning and scoping; (2) multi-modal evidence collection through NLP for document analysis, CV for physical control verification (supported by inspection robots and drones), and LLM-based conversational AI for interview; (3) ML-based evidence processing and triangulation; (4) confidence-weighted finding classification using Explainable AI; and (5) human-in-the-loop validation. The framework explicitly maps each module to the 93 controls of Annex A of ISO/IEC 27001:2022 and to the audit phases mandated by ISO 19011. We further propose a set of testable propositions, evaluation metrics, and ethical considerations that ground the framework in both academic rigour and practical deployability.

Abstract: Cities are in the middle of a parking transition. Minimum parking requirements are being reduced or eliminated, curbs are being repriced, and the goal of planning is shifting from supplying more parking to making better use of the parking that already exists. Yet most parking analytics still answer a question that this transition has retired: where should we build more? We argue that the distinctive value of agentic AI in parking is not better prediction of where to build, but the ability to expose contradictions that conventional workflows suppress—when demand says build but policy says restrain; when inherited rules say comply but theory says question; when market logic says maximize but equity says redistribute; and when stated public frustration says “parking crisis” but utilization data say the supply is ample and mispriced. Parking planning should be reconceptualized as a dynamic, theory-grounded, policy-constrained, human-supervised decision process, organized around a loop between parking theory, parking policy, urban data, agent reasoning, human deliberation, and policy revision—and ultimately answering a political question: what kind of city do we want to be? Under this view, an agentic parking system must be able to recommend shared parking, existing-stock reuse, curb and price reform, and deliberate non-construction, not only new supply. Using the Phoenix Parking Lot Planner as a critical demonstration—critical because its current weighted-factor scoring is precisely the kind of reasoning the proposed loop is meant to transcend—we outline a research agenda and five evaluation standards: contradiction detection, intervention comparison, justification quality, restraint capability, and policy traceability. Parking, precisely because it is measurable, theory-rich, policy-contested, and intervention-ready, may be the most realistic near-term testbed for agentic urban planning.

Abstract: The big data revolution transformed how we think of data analytics in many ways. Critical amongst them are the somewhat interconnected ideas of volunteered geographic information, crowdsourcing, and the big data property of variety. The robust literature concerning conceptual neighborhood graphs in two of these cases considers objects whose datatypes are held stable between the relations under consideration. This, however, is a limiting factor in these three application spaces due to the unknown form that data will take. This paper considers two avenues for the conceptual neighborhood graph to take as directions for future research: discretization conceptual neighborhood graphs (changing between corresponding vector and raster spaces) and cartographic generalization conceptual neighborhood graphs (changing the form of the objects in question). This paper provides insights as to what considerations should be considered when embarking upon this idea and demonstrates these concepts applied to prior conceptual neighborhood graphs.

Abstract: Digital trust in online interactions is commonly established through mechanisms such as decentralized identifiers (DIDs), verifiable credentials (VCs), and digital wallets. While these technologies ensure the correctness of individual components, they do not guarantee that an interaction as a whole is trustworthy. This limitation arises because real-world interactions consist of sequences of dependent steps, where inconsistencies may occur even when each step is locally valid. In this paper, we introduce the concept of executable trust, which models trust as a verifiable property of execution across interaction steps. We formalize interactions as sequences of TrustEvidence objects that capture both step-level validity and cross-step dependencies. Based on this model, we demonstrate that step-level correctness is insufficient to guarantee interaction-level trust, and we derive a minimal and sufficient condition for establishing end-to-end trust through composable verification and consistency constraints. We further present the Executable Trust Architecture (ETA), which operationalizes the proposed model through components for evidence generation, constraint enforcement, secure communication, and auditability. The feasibility and effectiveness of the approach are validated through scenario-based evaluations covering key trust properties, including authenticity, integrity, privacy, and accountability. The proposed approach provides a systematic foundation for verifying trust in complex digital interactions and supports the design of systems in which trust can be explicitly enforced, evaluated, and audited at runtime.

Abstract: A smallholder farmer in rural Karnataka spots something wrong with her tomato crop. She photographs the leaf, but the nearest agronomist is fifty kilometres away and charges fees she cannot afford. AgriAdvisor Pro is the system we built to close that gap. It pairs a fine-tuned EfficientNet-B2 classifier (97.88% accuracy, 65 classes) with Google Gemini 2.0 Flash for language generation, and stitches them together through a parallel orchestration layer running on Python’s asyncio. In concrete terms, this means advisory turnaround went from about 18 seconds down to 4.2—a difference that matters when your connection drops every few minutes. A FAISS-backed RAG pipeline ties each recommendation to verified regional documents rather than letting the model guess. We tested the system across 127 farms over one kharif season and saw a 34% drop in preventable crop losses along with less indiscriminate pesticide spraying. One season in one state is hardly definitive, and we are aware of that limitation. But even these preliminary numbers hint that designing around the farmer’s real constraints—patchy bandwidth, regional languages, limited digital literacy—can turn AI from a lab curiosity into something genuinely useful on the ground.

Abstract: Synthetic media, specifically AI-generated deepfakes, pose a growing threat to digital trust. As generation techniques improve, distinguishing authentic media from manipulations becomes increasingly difficult. This study presents a lightweight detection framework based on EfficientNet-B2, designed to balance computational efficiency with high forensic accuracy. Instead of retraining the entire network, we introduce a two-stage fine-tuning protocol. Initially, the backbone remains frozen while we train a custom classification head. Subsequently, we unfreeze the upper architectural blocks (Blocks 5 and 6) for specialized refinement using a reduced learning rate. This strategy preserves the general visual priors learned from ImageNet while adapting the model to the specific textural artifacts of deepfakes. We evaluated the system on a 140,000-image benchmark containing real FFHQ faces and StyleGAN outputs. On a hold-out test set of 10,905 images, the model achieved an AUC of 0.9624 and an overall accuracy of 88%. Notably, the model demonstrates a precision of 94% for the "fake" class, minimizing false accusations against real users. The training evolution highlights the efficacy of our approach: validation AUC jumped from 0.88 to 0.97 immediately upon unfreezing the deeper layers, eventually peaking near 0.995. These results suggest that targeted, layer-wise tuning allows smaller architectures to outperform traditional full-network transfer learning approaches.

Abstract: Aligned large language models (LLMs) often react very differently to the same jailbreak prompt: one model may refuse, another may partially comply, and a third may produce unsafe content. This variability suggests that jailbreak vulnerability is not determined by a single factor. Instead, it likely emerges from the interaction of backbone architecture, tokenization, prompt-template structure, post-training alignment, and internal representation-level mechanisms governing refusal and compliance. This concept paper argues that cross-model jailbreak variability should be studied as a mechanistic problem rather than only a benchmarking problem. Drawing on prior work on safety-training failure modes, optimization-based jailbreaks, shallow safety alignment, prompt-template effects, refusal directions, attention manipulation, and token-position sensitivity, this paper proposes a unified research agenda for explaining why aligned LLMs exhibit different internal responses to the same jailbreak prompt. The central thesis is that architecture matters, but many practically important differences arise from post training alignment and from how refusal and helpfulness are represented and routed internally.The paper formulates testable hypotheses, proposes an experimental framework spanning models such as Llama-2-Chat, Vicuna, and Mistral-Instruct, and outlines a methodology combining attack evaluation with attention analysis, hidden-state analysis, refusal-direction probing, tokenizer analysis, and causal interventions. The goal is to move from measuring jailbreak success toward understanding the internal mechanisms that produce it.

Abstract: Endpoint protection systems increasingly rely on a combination of signature-based and behavioral detection mechanisms, yet their robustness under systematic code transformation remains insufficiently understood. This paper presents a multi-category evaluation of endpoint detection robustness under automated, semantic-preserving code transformations across diverse execution variants. We introduce TransForge, a generalized transformation framework designed to generate functionally equivalent execution variants for controlled robustness assessment across heterogeneous artifact categories and programming environments. Building on our prior work, ShellForge, which focused on a single artifact class, TransForge extends this approach to support multi-category analysis through a modular transformation pipeline and an evolutionary strategy that enables non-deterministic variant generation. Using a dataset of 75 base samples spanning six execution categories and four programming languages, we conduct controlled experiments to evaluate how endpoint detection systems respond to systematically generated variants under consistent conditions. The findings reveal quantifiable variability in detection responses across categories and transformation strategies, highlighting coverage gaps in both signature-based and behavioral detection pipelines when faced with semantic-preserving transformations. This work motivates the development of robustness-aware evaluation frameworks and detection pipelines that leverage behavioral correlation and adaptive analysis beyond static signature matching.

Abstract: Large Language Models (LLMs) show promise in automated software engineering, yet their guarantee of correctness is frequently undermined by erroneous or hallucinated code. To enforce model honesty, formal verification requires LLMs to synthesize implementation logic alongside formal specifications that are subsequently proven correct by a mathematical verifier. However, the transition from informal natural language to precise formal specification remains an arduous task. Our work addresses this by providing the NaturalLanguage2VerifiedCode (NL2VC)-60 dataset: a collection of 60 complex algorithmic problems. We evaluate 11 randomly selected problem sets across seven open-weight LLMs using a tiered prompting strategy: contextless prompts, signature prompts providing structural anchors, and self-healing prompts utilizing iterative feedback from the Dafny verifier. To address vacuous verification, where models satisfy verifiers with trivial specifications, we integrate the uDebug platform to ensure functional validation. Our results show that while contextless prompting leads to near-universal failure, structural signatures and iterative self-healing facilitate a dramatic performance turnaround. Specifically, Gemma 4-31B achieved a 90.91% verification success rate, while GPT-OSS 120B rose from zero to 81.82% success with signature-guided feedback. These findings indicate that formal verification is now attainable for open-weight LLMs, which serve as effective apprentices for synthesizing complex annotations and facilitating high-assurance software development.

Abstract: Settlement discrepancies in multi-party electricity trading are difficult to localize because final outcomes are produced by multi-stage pipelines that combine heterogeneous data, rule versions, parameters, and execution contexts across organizational boundaries. In such settings, numerical reconciliation is not enough: investigators must identify where divergence entered the pipeline and support that judgment with evidence that can be checked independently. We formulate discrepancy localization as an auditable inference problem and introduce RootTrace, a trusted path-traceability method for this setting. Settlement processing is represented as an evidence graph over versioned artifacts and explicit events; RootTrace uses backward tracing and version-difference tracing to derive a suspect set of stages and/or artifact versions. The same pipeline exports a verifiable evidence bundle that preserves the trace used in localization. To support accountability under an explicit threat model, RootTrace includes trusted recording and verification procedures for tamper-evident capture and minimal-disclosure bundle export. On a semi-synthetic benchmark with ten independent replications, RootTrace achieves mean Top-1/Top-5/MRR of 0.47/0.83/0.60, compared with Top-1 0.17 and 0.04 for representative rule-based and stage-order baselines; exported bundles verify cleanly with full detection of the configured tamper classes (mean verification latency below 5ms on our grid); and an eight-hour unattended stress run completes over 1.3×10⁶ iterations without runtime failure.

Abstract: Attention mechanisms have achieved remarkable success in language models and have since been widely adopted in vision, speech, and multimodal learning. This trend has extended to graph learning, where attention-based models such as Graph Attention Networks (GAT) and Graph Transformers are now prevalent. \textbf{This position paper argues that attention mechanisms may not be as beneficial for graph node classification as commonly believed.} Through systematic ablation studies, we find that attention often provides negligible or even detrimental gains compared to simpler alternatives, with the only notable exception being graphs whose node features are language word embeddings. This suggests that the benefit of attention is largely limited outside language-related applications. We examine attention at three scales: 1-hop (GAT-style), Inception-style, and global mechanisms. We further analyze potential explanations for these results, including the limitations of gradient-based optimization and the fundamental differences between language and graph. Overall, these findings suggest that the prevailing enthusiasm for attention in graph node classification may be overstated, motivating a more critical and evidence-driven re-evaluation of its adoption. The code for all experiments is available at https://github.com/Qin87/ScaleNet/tree/July25.

Abstract: Whole-slide imaging has transformed histopathology into a data-intensive domain, with current approaches dominated by end-to-end deep learning that encode morphology implicitly within latent representations. This limits interpretability, reproducibility, and cross-dataset generalization. This review positions histomics as an intermediate phenotype representation layer that maps histological images to structured, multi-scale descriptors of tissue morphology, spatial organization, and architectural context. A unified taxonomy of histomic features across biological scales is presented, along with an analysis of artificial intelligence frameworks spanning classical machine learning, deep learning, weakly supervised learning, and multimodal integration. The review presents core failure modes in histomic pipelines, including segmentation dependence, feature instability, and domain shift, and examines their impact on robustness and generalization. Emerging trends in representation learning and multimodal modeling are analyzed in the context of phenotype-centric inference. Overall, this work reframes histomics as a representation-driven paradigm and outlines directions for developing stable, interpretable, and generalizable computational pathology systems.

Abstract: The transition from smart to intelligent cities allows for the deployment and management of information and communication technologies in the urban context to be driven by holistic sustainability requirements rather than technical ones such as feasibility and fragmented, siloed operational patterns. This work proposes a multi-dimensional decision-making framework to manage a smart-intelligent city as an urban Cyber-Physical System across environmental, economic, and social sustainability pillars, metrics and their tradeoffs. A methodology based on Deep Reinforcement Learning and reward-shaping mechanisms is proposed to represent and assess sustainability pillar dependencies and their interplay. A case study on a Low-Power Wide-Area Network planning, deployment and management in a Sicilian municipality has been developed to demonstrate the effectiveness of the proposed approach in dealing with the dynamics and the non-linear dependencies of the sustainability pillars. The results thus obtained provide a blueprint for urban planners to develop sustainable, resilient, cost-effective, and environmentally friendly smart-intelligent city frameworks.

Abstract: Existing travel planning systems lack user participation in itinerary scoring and apply coarse, binary weather treatment that risks excluding high-quality outdoor attractions under mild precipitation. This paper presents a multi-objective genetic algorithm (GA)-based itinerary planning system that addresses both limitations. The system incorporates a weather-adaptive POI scoring framework mapping nine weather conditions to three strategies, applying intensity-proportional rain penalties and a geographic flexibility bonus scaled by local indoor alternative density. User preferences are encoded via three integer sliders whose normalised values directly set the GA fitness weights for POI quality, traveling efficiency, and preference satisfaction. The system is evaluated on 144 attractions in Macao. Results show that outdoor POI representation decreases proportionally with precipitation intensity across all nine weather conditions and is substantially suppressed under official extreme weather alerts, while itinerary quality is preserved through the flexibility bonus. Slider adjustment experiments confirm that amplifying each weight produces statistically consistent, direction-correct improvements in its target sub-objective without degrading the others. These findings validate the functional independence of the three-objective fitness formulation and demonstrate that graduated weather treatment and direct user weight control together yield a more responsive and robust itinerary planning system.