Computer Science and Mathematics

Abstract: Self-attention is fundamental to Transformer architectures, but its computational and memory requirements grow quadratically with sequence length. This paper presents a structured narrative review of computationally efficient self-attention algorithms. It establishes the mathematical foundations of standard self-attention and examines dense, sparse/local, low-rank projection, landmark/Nyström, kernel/linear, memory-efficient exact, and hierarchical mechanisms. The review compares these approaches in terms of arithmetic complexity, memory requirements, sequence-length scaling, approximation, connectivity, I/O behaviour, implementation overhead, and hardware dependence. It distinguishes algorithmic complexity reduction from implementation-level optimization and clarifies the conditions under which complexity is linear in sequence length. Analytical crossover conditions show how sequence length, representation dimension, mechanism-specific parameters, and implementation overhead influence practical advantage. The review demonstrates that efficient attention cannot be characterized solely as quadratic versus linear complexity and identifies research gaps in multidimensional complexity modelling, approximation-error trade-offs, I/O complexity, long-range dependency preservation, Pareto efficiency, and hardware-aware analysis.

Abstract: Large language model (LLM) agents are no longer deployed as bare models. They run inside an execution harness, the control layer that assembles context, exposes tools, manages memory and state, schedules retries, and verifies intermediate results. Because the harness determines what the model sees and what its outputs are allowed to do, observed agent behavior is jointly produced by the model and the harness, yet prevailing benchmarks attribute it to the model alone. We study the problem of harness evaluation: holding the model, task distribution, environment, budget, and evaluator fixed, how can the contribution of the harness be isolated, and what does it mean for a harness to be trustworthy? We formalize the harness as a trajectory-generating policy and define the harness effect as a paired, metric-indexed contrast. We then develop measurement principles along seven dimensions—outcome correctness, quality, efficiency, reliability and reproducibility, robustness, safety, and fairness—each targeting the specific mechanisms through which a harness influences agent behavior, and grounded in established results from statistical evaluation, efficient inference, certified robustness, and fair learning. A set of elementary propositions clarifies when paired designs reduce variance, how approximate verification bounds correctness loss, why marginal subgroup gaps can mislead, and how differential evaluator bias contaminates harness comparisons. We conclude with a reference protocol and argue that harness quality should be reported as a Pareto profile rather than a single leaderboard score.

Abstract: We introduce an explicitly axiomatized mathematical model called the density neutrosophic operator. The construction is motivated by the quantum-mechanical density operator and by the neutrosophic distinction between truth, indeterminacy, and falsity. We emphasize that the terminology is proposed here and does not presuppose an established standard definition. Let \( \mathcal{H} \) be a finite-dimensional complex Hilbert space. A density neutrosophic operator is defined as a triple \( \mathbf{\rho} =(\rho_T,\rho_I,\rho_F) \) of positive semidefinite operators whose sum is a quantum density operator. The components represent three nonnegative operator contributions, interpreted respectively as truth, indeterminacy, and falsity contributions within this mathematical model. We establish existence criteria, spectral properties, convexity, and a complete characterization of pure states. In particular, the aggregate state is pure if and only if all three component operators are nonnegative scalar multiples of one common rank-one projection. Mixed states are characterized by aggregate rank greater than one, and their decompositions are examined through examples and counterexamples. We also study measurement probabilities, unitary covariance, partial traces, and extensions to completely positive trace-preserving maps. The results identify which properties follow from the proposed axioms and which additional interpretations would require further physical or mathematical assumptions.

Abstract: Improper waste disposal is a major environmental problem because dry, wet, and metal waste are frequently mixed together. Traditional sensor-based waste segregation systems can work effectively when waste items are placed individually, but their performance can be limited when different types of waste are mixed. This research proposes an IoT-Based Smart Waste Management System using YOLO Computer Vision for automatic waste segregation. A camera is used to capture waste images and the YOLOv8 model is used to identify and classify waste. The proposed system integrates Computer Vision with an IoT-based hardware prototype consisting of Arduino UNO, IR sensor, Raindrop moisture sensor, servo motor, stepper motor, and ESP8266 Wi-Fi module. The detected waste category is used to control the segregation mechanism. The research evaluates the proposed system using Computer Vision and system-level performance metrics. The objective is to improve the accuracy, speed, and automation of waste segregation while reducing dependence on manual sorting.

Abstract: Biomedical Knowledge Graphs (KGs) such as PrimeKG and Hetionet offer vast structured repositories for clinical discovery; however, conventional Graph Neural Networks (GNNs) suffer from black-box opacity, while standard Retrieval-Augmented Generation (RAG) paradigms face catastrophic hallucinations and biological implausibility. To resolve this critical trade-off, this paper proposes BioAgent-RAG, a multi-agent orchestrated retrieval-augmented generation framework tailored for complex biomedical reasoning and transparent graph traversal. By integrating autonomous sub-agents for query decomposition, path-based KG grounding, and cross-verification, BioAgent-RAG systematically enhances biological plausibility and reduces reasoning hallucinations. Comprehensive empirical evaluations demonstrate that BioAgent-RAG substantially outperforms existing competitive baselines, RD2-KG, KGiA, and RAG-LLMs. Specifically, on the PrimeKG benchmark, it achieves an ROC-AUC of 0.952, PR-AUC of 0.814, MRR of 0.637, and superior Recall@10 and Recall@50 of 0.715 and 0.868, respectively (p < 0.01). Moreover, evaluated on biological plausibility, BioAgent-RAG yields a Biomedical Factuality Score (BFS) of 0.887 on PrimeKG (0.894 on Hetionet) and a Path Reasoning Score (PRS) of 0.841 (0.862 on Hetionet), demonstrating strong preliminary correlation (0.912) with domain-expert evaluations. These findings suggest that BioAgent-RAG serves as a robust, explainable computational screening tool to prioritize drug repurposing candidates prior to in vitro validation.

Abstract: Standardized cognitive and functional assessments, such as those in the National Alzheimer’s Coordinating Center (NACC) Uniform Data Set (UDS), are central to detecting and monitoring Alzheimer’s disease, yet their administration and scoring depend on trained examiners who record and score responses by hand. We present a modular speech-processing pipeline that captures and structures spoken responses from UDS assessments administered by a clinician or by the Zeus2Q humanoid robot. The pipeline converts recordings to 16-kHz mono audio, segments speakers with the pyannote.audio 3.1 diarization pipeline, transcribes each speaker turn with Whisper medium.en, optionally personalized by participant-level fine-tuning, and emits a speaker-attributed, time-stamped JSON record aligned with UDS items for downstream scoring. The pipeline runs end to end on Google Colab and on an NVIDIA H200 GPU cluster. In a pilot on volunteer recordings of UDS-4 Form B7, the NACC Functional Assessment Scale, a fine-tuned Whisper model achieved a normalized word error rate (WER) of 25.7% and a character error rate of 18.1% on an evaluation recording. Error analysis showed that a single deleted response would produce at least two incorrect item codes if responses were assigned to items in order of occurrence, so WER alone does not capture clinical utility. Continued training and a single-example run drove the training loss to near zero within 40 steps, which reflects memorization rather than generalization and highlights the overfitting risk of few-sample personalization. We specify the evaluation protocol for a planned 200-participant study: diarization error rate, held-out WER against a zero-shot baseline, and item-level agreement with clinician scoring.

Abstract: Large language models (LLMs) can produce plausible but incorrect content, a risk that is unacceptable when public institutions disclose scientific and legal information. This study presents and evaluates a hybrid agentic retrieval-augmented generation (RAG) system that automates first-level responses to citizen information requests at the Marine Institute of Peru (IMARPE). A baseline analysis of 196 requests showed a mean response time of 12 business days, above the 10-day legal limit, and identified 26.9% of them as routine, data-extractable queries. The system, developed under CRISP-DM, uses LangGraph to route each query to a LightRAG knowledge graph (Neo4j) built from institutional yearbooks, to a Text-to-SQL agent that queries the IMARSIS relational database through the Model Context Protocol (MCP), or to both, with Gemini 2.5 Flash as the foundation model and a referral protocol for out-of-scope requests. Five institutional domain experts rated 30 test cases, with mean scores of 4.45/5 for factual accuracy, 4.10/5 for completeness, and 4.57/5 for clarity. Mean end-to-end software latency was 20.43 s (maximum 49.20 s), and 30 pilot users rated usability at 80.42 on the System Usability Scale (Cronbach’s α = 0.96). These preliminary, single-site results indicate that combining graph-based and SQL-based grounding is feasible for routine citizen-information services; comparison with simpler baselines remains pending.

Abstract: Background/Objectives: Primary, idiopathic, impaired spermatogenesis (IS) affects approximately 2% of men and can manifest as oligozoospermia or azoospermia. Methods: In order to gain insights into the genetic mechanisms involved, a systematic analysis of historical microarray data from testes with IS was conducted by first constructing a global map of signalling networks involving genes differentially expressed between IS and control testes (CT). The centrally located genes and their interacting genes were then extracted to create a sub-network representing the pathways active in IS. Results: Two routes to germ cell death or cycle arrest were revealed: either via receptor tyrosine kinases or integrin β (ITGB) genes; the latter being a novel finding. Another analysis, using leave-one-out machine-learning validation, was performed and also showed that ITGB expression distinguished IS from CT. Computational simulation confirmed the ITGB protein family regulates a majority of the other proteins in its pathway whether the expression level was reduced or increased, indicating ITGB could be a key protein in IS. Conclusions: These computational analyses identify ITGB, particularly ITGB1, as a potential molecular marker associated with idiopathic impaired spermatogenesis and provide insights into signalling mechanisms underlying the condition.

Abstract: Retrieval-Augmented Generation (RAG) enhances text-to-image diffusion models by grounding generation in retrieved visual exemplars, but recent studies reveal that multimodal retrieval pipelines are highly vulnerable to poisoning attacks. When adversaries corrupt the retrieval database, semantically mismatched exemplars such as retrieving images of cats for prompts requesting dogs can mislead diffusion models into generating incorrect or misleading outputs. We identify this failure mode as a breakdown of semantic coherence between the text prompt and retrieved visual context. To address this issue, we propose a score-based semantic coherence refinement module that explicitly evaluates prompt-image consistency, refines misaligned prompt components, and re-retrieves corrected exemplars prior to diffusion. Acting as a multimodal feedback loop, the proposed method prevents poisoned retrieval from propagating semantic errors into the generative process. Extensive experiments demonstrate that our approach significantly improves semantic correctness, alignment, and robustness under both clean and poisoned retrieval settings, establishing an effective and principled defense for Vision RAG-augmented diffusion models.

Abstract: Reinforcement learning (RL) has emerged as a powerful paradigm for addressing sequential decision-making challenges across the software development life cycle (SDLC). To understand how this paradigm is being operationalized in practice, this systematic literature review examines 85 studies published between January 2020 and March 2026 that apply RL to the seven core SDLC phases: planning, analysis, design, development, testing, deployment, and maintenance. We analyze publication trends and venues, map RL applications to specific SDLC tasks, catalog the RL algorithms employed, and provide a taxonomy of reward modeling strategies and their parameters.Our findings reveal a rapid growth of RL-for-SDLC research since 2022, with the testing phase being the most extensively studied (56.47% of studies) while other phases show steady expansion. Reward designs have evolved from sparse binary outcomes (e.g., test-pass/fail) to dense, multi-objective signals that incorporate execution semantics, coverage metrics, intrinsic curiosity, formal verification feedback, and human preferences. We highlight common reward patterns, their influence on learning behavior, and trade-offs that frequently lead to reward hacking or unstable training in software environments. This survey provides researchers and practitioners with a structured overview of how RL is operationalized across the SDLC, emphasizing reward engineering as a critical factor in agent effectiveness. We conclude by outlining promising directions for future research, including broader lifecycle coverage and the development of more robust reward mechanisms for production-ready software engineering automation.

Abstract: With 60 years of research, knowledge-based system has travelled a long journey from early small-scale rule-based programs to modern super-large-scale knowledge graphs. Several multithreading technical stages can be broadly categorized: early history of rule-based system, knowledge engineering, semantic network, semantic web, ontology engineering, knowledge graph, and recent advances driven by machine learning. In the past couple years, the whole area of artificial intelligence including knowledge-based system has been significantly influenced by large language models, which dramatically increases the understanding and utilization of knowledge by increasing to an extremely large scale. Considering the long-standing intrinsic problems of large language models such as hallucination and lack of explainability, the challenge of how to build reliable large-scale knowledge-based systems in an effective and economical way still exists. From the overview of the developing history, we can see a promising research direction of combining large language models with those traditional knowledge-based frameworks.

Abstract: Computational research increasingly depends on custom software, yet much of this software begins as project-specific analysis code written to answer one immediate scientific question. Such code can be appropriate and productive in its original context, but it often becomes fragile when it is reused, shared, cited, maintained, or extended by others. In this perspective, we frame the transition from project-specific analysis code to sustainable scientific software as a deliberate development process rather than an automatic consequence of increasing code size, algorithmic sophistication, or wider use. This transition depends on development decisions that make computational work reproducible, inspectable, reusable, and creditable. We propose seven principles for sustainable research software: modular design, version control, automated testing, documentation, public availability and open licensing, persistent citation, and transparent development. These principles are most effective when considered from the beginning of a project, but they can also be introduced retrospectively as part of software maturation. We discuss how the meaning of code changes when it becomes shared scientific infrastructure, why AI-assisted and agentic coding require accountable development practices, and why research software should be understood as a core topic in research data management. Finally, we illustrate how the principles can be implemented in practice through a pseudo-code example. Sustainable scientific software is not merely cleaner code. It is research infrastructure that can be inspected, reused, tested, cited, and maintained.

Abstract: Social bot detectors can perform well on a single benchmark but degrade when bot populations, feature availability, annotation practices, or graph structure change. This study evaluates multi-benchmark multimodal graph pretraining for cross-benchmark social bot detection under distribution shift. A shared encoder integrates profile, textual, behavioral, temporal, and relational signals and is assessed on TwiBot-20, Twi-Bot-22, Cresci-2015, and Multi-Relational Graph-Based Twitter Account Detection Benchmark (MGTAB) using within-benchmark classification, direct source-to-target transfer, leave-one-benchmark-out pretraining, and sequential adaptation. Mean within-benchmark Macro-F1 was 94.13%, whereas direct transfer produced generalization gaps of 2.78–16.13 percentage points. Pretraining on three benchmarks improved mean held-out F1 from 90.54% when training from scratch to 93.80% after fine-tuning. After sequential exposure to all four benchmarks, mean F1 remained 92.89%, with retrospective losses of 0.66–2.11 points on previously seen tasks. These results show that multi-benchmark pretraining improves target adaptation but does not eliminate distribution shift. The study separates in-domain accuracy, transferability, and retention, providing a focused evaluation framework for robust social bot detection.

Abstract: Finite-state machines (FSMs) are widely used to implement control logic in digital systems and are therefore important targets for intellectual property protection and hard ware obfuscation. Although FSM obfuscation aims to conceal the underlying control behavior, its implementation may leave measurable characteristics that distinguish obfuscated FSM (OFSM) state registers from other sequential elements. This work investigates whether OFSM state registers protected us ing the ProtectIP state-space obfuscation methodology can be localized using only timing and switching-activity information, without relying on register names or explicit structural features. Specifically, setup timing, hold timing, and switching activity are used to construct a numerical profile for each flip-flop. LOCATE is evaluated across six benchmark designs containing 4,743 flip flops, including 331 OFSM state registers. Under within-design five-fold validation, the approach identifies 281 of the 331 OFSM registers with only 14 false positives, achieving 95.3% precision, 84.9% recall, and an F1 score of 89.8%. More importantly, leave-one-design-out cross-design evaluation, in which the target circuit is completely excluded from training, achieves 87.2% precision, 80.4% recall, and an F1 score of 83.6%. Feature ablation analysis further shows that setup and hold timing provide the strongest discrimination, while switching activity contributes complementary information that improves overall localization performance. These results demonstrate that FSM obfuscation can leave transferable implementation-level finger prints and highlight timing and switching activity as potential sources of information leakage for localizing protected state registers in previously unseen obfuscated designs.

Abstract: This study investigates exact algorithms for solving the Travelling Salesman Problem (TSP), formulated as the problem of finding a minimum Hamiltonian cycle in a complete weighted undirected graph. Two recursive algorithms are presented and analyzed. The first algorithm, denoted RecursiveHamiltonianCycle (RHC), performs an exhaustive recursive search of the solution space, while the second algorithm, denoted PrunedHamiltonianCycleSearch (PHCS), applies an early branch-pruning mechanism to eliminate non-promising partial solutions. The main objective of this research is to evaluate the influence of recursive branch pruning and initial upper-bound selection on the computational efficiency of exact TSP search procedures. A series of computational experiments was conducted using complete graphs containing between 12 and 21 vertices. The obtained results show that both algorithms always produce identical optimal Hamiltonian cycles. However, the PHCS algorithm requires significantly fewer recursive calls and substantially less execution time. Additional experiments demonstrate that the quality of the initial upper bound has a substantial impact on pruning efficiency. A nearest-neighbor heuristic was therefore used to generate high-quality initial solutions, resulting in further reductions in the search space. The results confirm that combining heuristic initialization with exact branch-pruning search significantly improves the practical applicability of exact TSP algorithms while preserving solution optimality.

Abstract: Multi-objective Next Release Problem experiments require the coordinated management of instance data, requirement interactions, optimisation configurations, solver execution, and result analysis. These activities are often distributed across independent scripts, libraries, and visualisation tools, making it difficult to preserve consistent problem semantics and traceability throughout the experimental workflow. This paper presents NRP-Exp, a modular open Python framework that connects instance provision, evaluator aggregation, interaction-aware preprocessing, multi-objective problem formulation, solver execution, Pareto-front assessment, and post-optimisation analysis through common representations and reusable headless components. Its ordered preprocessing pipeline transforms descriptive NRP instances into traceable optimisation-ready variants by resolving coupling components, remapping implication relationships, processing value-based interactions, and constructing alternative branches for mutual exclusions. The framework integrates six heterogeneous optimisation algorithms through a common solver contract while preserving their applicability conditions. Validation combines an end-to-end reference workflow, six NRP formulation profiles containing between 12 and 100 requirements, and a controlled linear case for solver portability. All formulation cases were loaded and transformed successfully, producing nine optimisation-ready variants; each selected case was executed through a compatible solver, and all six solver implementations returned feasible results for the common controlled formulation through the shared solution and Pareto-front representations. These results support NRP-Exp as an integrated and traceable infrastructure for structurally repeatable and configurable NRP experiments, while highlighting that strict experimental reproducibility with stochastic solvers requires explicit random-seed management.

Abstract: Prototyping is an important component of user interface development, during which key decisions are made regarding methods of information presentation and the organization of user interaction. Assessing inclusivity at this stage makes it possible to account for relevant factors before the final implementation of the interface and to adjust design decisions accordingly. The goal of the study is to develop a model for fuzzy multi-criteria assessment of the user interface inclusivity at the prototyping stage, taking into account the relative importance of factors and the preference degree of alternatives. The relative priority of factors is established by the pairwise comparison method using the Saaty scale. To demonstrate the application of the model, three different options for the interface solution at the prototyping stage are considered. For their comparison, fuzzy preference relations are formed for each factor, their intersection and weighted convolution are performed. Sets of non-dominated alternatives and the resulting membership function are determined. A structural-functional model and pseudocode of the assessment algorithm are developed. The proposed approach makes it possible to quantitatively compare prototypes based on a set of inclusivity factors and determine the optimal interface solution.

Abstract: The increasing integration of artificial intelligence (AI) into education creates opportunities to support learning, but its educational value depends not only on technological capabilities but also on human–AI interaction. This issue is particularly relevant in interpreter training, where learning involves high cognitive demands, contextual decision-making, and time-sensitive activities. This study investigates how Human-Centred Artificial Intelligence (HCAI) principles can guide the design of AI-assisted educational environments for interpreter training. A Design Science Research (DSR) methodology was adopted to develop and evaluate an AI-assisted learning environment integrating AI capabilities across the interpreter training cycle. Drawing on literature on Computer-Assisted Interpreting, HCAI, human–AI interaction, technology acceptance, and interpreting cognition, five design principles were derived: learner agency and human control, transparency and calibrated trust, protection of cognitive resources, contextualised assistance, and AI as a support for learning. These principles were operationalised in the artefact and evaluated through heuristic evaluation and a technology acceptance study involving 42 professional interpreters. Results show that perceived usefulness alone is insufficient to explain technology acceptance, with trust, reliability, ease of use, latency, workflow integration, and cognitive demands also shaping adoption. The study contributes a human-centred framework for aligning AI capabilities with learners’ needs and educational objectives.

Abstract: Deep-learning track reconstruction methods such as hepattn and HEPTv2 perform strongly on the TrackML benchmark, but each represents detector hits from a single, architecturally bounded viewpoint. This work asks whether a channel-boosted multi-agent architecture can recover more particle tracks than either expert model achieves alone. The proposed system, CB-HMAS (Channel-Boosted Hierarchical MultiAgent System), combines a newly designed, physics-informed Relational Anchor Agent with two frozen expert peer channels (hepattn and HEPTv2) through a hyper-attention Fusion agent. A reinforcement-learning Topology Controller adjusts each peer channel’s contribution, and a meta-learning optimizer provides per-event adaptation. We evaluated the system on 210 held-out TrackML events (IDs 3000–3210, drawn from a separate data folder with zero overlap with the 1000–2300 training range). Across nearly the full range of similarity thresholds tested, the fused system exceeds both individual peer channels; at threshold 0.9 it reaches 76.0% efficiency, against 42.0% for hepattn alone and 12.2% for HEPTv2 alone. An embedding-capacity ablation identified representational capacity as a limiting factor: increasing the shared Anchor/Fusion embedding dimension from 32 to 64 raised peak efficiency from 45.7% to 76.0% under an otherwise identical configuration. Across five independent training runs with different random seeds, efficiency at threshold 0.9 is stable at 75.6% ± 2.4 percentage points, and the fused system outperforms both individual peer channels for every seed tested. These held-out results indicate that channel boosting recovers track structure that neither expert channel captures alone, and that the advantage is consistent rather than the product of a single favorable run.

Abstract: Cyber Threat Intelligence (CTI) sharing is critical for collaborative defense, yet organizations struggle to assess the reliability and quality of shared information in standardized formats like STIX 2.1. This trust deficit stems from difficulties with traceability, data integrity verification, and distinguishing high-fidelity threats from low-quality alerts, undermining the effectiveness of collective security efforts. Prior work introduced BlockCTI, an architectural solution that combines a permissioned blockchain with a Proof of Quality (PoQ) consensus mechanism grounded in NATO Admiralty Scale criteria to address these challenges. While theoretically sound, the practical viability of this design remained unvalidated. This paper presents the implementation and empirical validation of BlockCTI, demonstrating that the proposed model translates effectively into a working system. We developed a functional prototype integrating blockchain-based incident logging with automated quality assessment and TAXII/MISP compatibility. Comprehensive experimental evaluation across five dimensions: scalability, validator sensitivity, resilient coordination, load variation, and quality robustness, validates the core theoretical predictions: the system exhibits predictable latency and throughput characteristics as demand increases, and the PoQ mechanism successfully differentiates among high, medium, and low-reliability incidents. These findings confirm that BlockCTI’s theoretical model is practical and deployable, offering a viable solution for trustworthy, auditable CTI sharing in collaborative government environments.