Abstract: Refactoring is essential for developing maintainable software. Using Large Language Models in software engineering is widespread, but compared to well-established domains such as code generation, reliable refactoring is still relatively underexplored. In this paper, we perform a broad analysis on the refactoring capabilities of small open-weight language models (SLMs) by evaluating 12 models on 3,453 Python programs. Our study focuses on the two defining aspects of refactoring: behavior preservation and code quality improvement. We evaluate these properties using unit tests and various code metrics. Across models ranging from 0.5B to 8B parameters, most models improve code quality. Larger models are more reliable, as they preserve behavior more consistently. Reasoning models often make more significant changes while refactoring. Allowing models to generate reasoning traces improves performance, but only for models larger than 4B. For smaller models, reasoning in fact reduces refactoring reliability. The difficulty of the underlying task affects refactoring performance, with more complex tasks associated with higher failure rates. Our results indicate that current open SLMs can support refactoring tasks, especially larger ones with reasoning capabilities, but they are best used with human oversight.

Abstract: Large language models (LLMs) face significant challenges in sustaining long-term memory for agentic applications due to limited context windows. To address this limitation, many work has proposed diverse memory mechanisms to support long-term, multi-turn interactions, leveraging different approaches tailored to distinct memory storage objects, such as KV caches. In this survey, we present a unified taxonomy that organizes memory systems for long-context scenarios by decoupling memory abstractions from model-specific inference and training methods. We categorize LLM memory into three primary paradigms: natural language tokens, intermediate representations and parameters. For each paradigm, we organize existing methods by three management stages, including memory construction, update, and query, so that long-context memory mechanisms can be described in a consistent way across system designs, with their implementation choices and constraints made explicit. Finally, we outline key research directions for long-context memory system design.

Abstract: This study integrated Item Response Theory (IRT) models with ordinal survey instruments to assess academic performance trajectories and identify multidimensional factors associated with academic achievement among first-semester leveling students (N=1,558 pre-test; N=1,676 post-test) at the Escuela Politécnica Nacional, Ecuador. A dual-component methodology was employed: (1) an 80-item ordinal survey measuring eight latent constructs (socioeconomic, academic, motivational, vocational, social integration, psychological/emotional, institutional, and biological/health factors), validated through Confirmatory Factor Analysis (CFI > 0.95, RMSEA < 0.06); and (2) structured diagnostic assessments in mathematics, physics, chemistry, geometry, and language, calibrated using three-parameter logistic (3PL) IRT models via Expected A Posteriori (EAP) estimation. Results demonstrated high internal consistency (r = 0.93 between IRT and raw scores), with mean IRT-scaled ability θ ̅ = 10.45 (SD = 3.51) on a 1–20 scale. Item parameters indicated adequate discrimination a ̅ = 1.92) and centered difficulty (b ̅ = 0.05), though 13.75% of items exhibited poor model fit (S-X² p < 0.01), concentrated in physics and chemistry domains. Factorial scores and performance outcomes were statistically contrasted against 24 categorical demographic variables, revealing differential performance patterns across student subgroups. This research provides validated psychometric instruments, reproducible IRT-LMS integration protocols, and empirical evidence supporting targeted interventions to strengthen university transition in resource-constrained contexts.

Abstract: Cloud-based large language models (LLMs) have demonstrated near-human performance in medical applications; however, their clinical deployment is constrained by concerns regarding patient privacy, data security, and network dependence. Locally deployable, open-weight LLMs may provide a privacy-preserving alternative for resource-limited or security-sensitive environments. We evaluated two families of locally deployed models, Google Gemma3 (1B, 4B, 12B, and 27B parameters; vision enabled in models since 4B) and GPT-OSS-20B, using 1,200 multiple-choice questions from the Taiwan Pulmonary Specialist Board Examinations (2013–2024), including 1,156 text-only and 44 text-and-image items across 26 categories. A cloud-based GPT-4 Turbo model served as a reference. Models were queried locally via Ollama. Accuracy was analyzed by year and category using repeated-measures ANOVA with Tukey-adjusted pairwise comparisons. GPT-OSS-20B achieved the highest overall accuracy (58–78 correct answers per 100 questions) and significantly outperformed all Gemma-3 variants (p &lt; 0.001), while Gemma3-27B ranked second. No statistically significant difference was observed between GPT-OSS-20B and GPT-4 Turbo after Tukey adjustment. Larger models showed improved accuracy but longer inference time. These findings suggest that selected open-weight LLMs deployed on-device can approach the performance of cloud-based models in structured medical examinations, with trade-offs between accuracy, modality support, and computational efficiency.

Abstract: Recent advances in large language models (LLMs) have made it feasible to use them as automated debugging tutors, but it remains unclear how much can be gained by moving from single-model tutors to multi-agent councils with separated roles. We study this question in an offline simulation on 200 debugging cases drawn from an online judge, spanning 20 problems split into course-style and contest-style challenge tracks. We compare four single-model tutors based on current frontier models with four councils that assign models to Architect, Skeptic, Secretary, Pedagogue, and Mentor roles and operate in both Blind and Guided modes. Single-model tutors achieve near-perfect repair on course problems but perform less reliably on challenge cases and often rewrite large portions of student code, show non-negligible false positive rates, and leak full or near-full solutions in a substantial share of hints. Councils designed around measured model strengths improve both technical and pedagogical behaviour. On the challenge track, the best council raises patch success by 12.2 percentage points over the best single tutor, while reducing false positives, shrinking median patch size, improving hint localisation, and cutting solution leakage in Blind mode from about one fifth of hints to under ten percent. Councils also exhibit higher stability across reruns and produce hints that two independent instructors consistently rate as more useful and better scaffolded. Guided mode, where internal components see a reference solution, yields further technical gains but introduces leakage risks that require prompt tightening and a sanitising Secretary to control the flow of ground truth. Additional trap experiments with poisoned reference solutions show a mix of resistance and fail-safe collapse rather than systematic poisoning of hints. These results indicate that orchestration and information flow are powerful levers and that well-designed councils can provide more reliable and pedagogically aligned debugging support than strong single-model tutors alone.

Abstract: Large language models (LLMs) require a significant redesign in solutions to preserve privacy in data-intensive applications due to their text-generation capabilities. Indeed, LLMs tend to memorize and emit private information when maliciously prompted. In this paper, we introduce Private Association Editing (PAE) as a novel defense approach for private data leakage. PAE is designed to effectively remove Personally Identifiable Information (PII) without retraining the model. Experimental results demonstrate the effectiveness of PAE with respect to alternative baseline methods. We believe PAE will serve as a critical tool in the ongoing effort to protect data privacy in LLMs, encouraging the development of safer models for real-world applications.

Abstract: This paper presents a generative AI frameworkfor producing structured symbolic sequences with fine-grainedexpressive control. The approach introduces a compact tokenrepresentation combined with phrase-aware latent alignment tosupport coherent generation across variable-length segments. Byintegrating sequence-level regularization directly into attention,the model balances structural consistency and diversity withoutrelying on explicit post-processing constraints. Empirical analysisshows that the method maintains stable distributional behavioracross expressive dimensions, highlighting its suitability forcontrollable symbolic generation tasks.

Abstract: This paper propose a formal framework for AIagents that unifies semantic reasoning with resource-aware con-trol. Agents act via sparse policies over structured semantic fields,bounded by entropy and sparsity budgets. We define a typedoperational semantics, prove soundness and stability, and derivea sparse free-energy objective with phase transitions. The calculusis categorically structured, maps to unistochastic dynamics, andcompiles to executable policies with verified runtime bounds.This yields a foundation for interpretable, thermodynamically-plausible agent design.

Abstract: It is a universally acknowledged heuristic of science that, all else being equal, a theory with fewer free parameters that explains more empirical data is superior. Yet, this intuitive preference is rarely formalized into a strict, operational objective function. This paper formally translates that heuristic into an invariant mathematical boundary condition from a strict Tarski Level-2 vantage point. We advance the Minimum Description Length (MDL) principle—grounded in Algorithmic Information Theory (AIT) and Computability Theory (CT)—not as a philosophical preference, but as the absolute, objective metric for evaluating cross-domain scientific progress. Because empirical science is formally defined as an inductive computational search over a strictly finite observation string (\( $\Sigma_t$ \)), genuine foundational advancement occurs if and only if the computable surrogate of total description length strictly decreases (\( $\Delta \hat{L} < 0$ \)).We establish a rigorous algorithmic boundary between descriptive Engineering Maps and the constructive Engine of reality. Because every unconstrained parameter added to a generative program carries an inescapable exponential penalty in algorithmic probability, the post-hoc addition of unobservable latent variables (NODF Inflation) mathematically guarantees theoretical degradation. Furthermore, by the strict laws of computability, theories relying on infinite-precision continuous mathematics evaluate to an infinite informational cost (\( $L=\infty$ \)) and are structurally disqualified from foundational ontology. We map how modern institutional topologies systematically evade this algorithmic metric through statistical thresholding (discarding high-information anomalies) and VC-dimension inflation (parameter patching). The burden of proof now rests on proposing a mathematically superior metric for scientific progress that does not rely on self-referential sociological consensus.

Abstract: Artificial Intelligence (AI) refers to systems designed to mimic human intelligence, enabling machines to perform tasks that typically require reasoning, learning, and decision-making. Today, AI is integrated into everyday life through technologies such as virtual assistants (e.g., Siri, Alexa, and Google Assistant), autonomous transportation systems, aviation technologies, gaming, and digital platforms. While AI has transformed multiple industries, healthcare has emerged as one of its most impactful domains, significantly enhancing medical imaging, disease diagnosis, treatment planning, and patient management. However, the clinical adoption of AI has been constrained by several persistent barriers, including limited computational resources, scarcity of high-quality annotated datasets, lack of interpretability, privacy concerns, regulatory ambiguity, and integration challenges within existing healthcare infrastructures. The emergence and rapid advancement of modern Deep Learning (DL) techniques helped address many of these challenges by enabling AI systems to analyze complex and high-dimensional healthcare data more effectively. Consequently, AI is increasingly leveraged to overcome traditional healthcare system constraints, improving diagnostic precision, workflow efficiency, and patient outcomes. Despite significant progress, unresolved technical, ethical,regulatory and organizational challenges necessitate a comprehensive evaluation of AI’s role in healthcare. This review discusses the evolution and applications of AI in healthcare, examines the limitations of traditional healthcare systems, explores how AI addresses these challenges, identifies current limitations of AI based approaches, and presents potential solutions to guide future advancements in AI applications within healthcare systems.

Abstract: Mobile-based cloud computing (MBCC) has become a paradigm shift that greatly en-hances the computing power of the mobile devices with limited resources by enabling them to access in-the-cloud resources of highly scalable infrastructures at considerable distance. Nevertheless, the constant and unregulated data transfer between mobile customers and remote cloud providers is deep bandwidth expenses, and at the same time, interfere with the overall system performance, address undesirable energy utili-zation, and sensitive information to security weaknesses. These are especially the problem in bandwidth-limited geographical locations or cost-aware enterprise settings in which the network usage is directly translated into expensive operational costs. This paper proposes a hierarchical, edge-aware architecture of the system in an approach to provisioning a holistic optimization of bandwidth usage as long as end-users remain robust in performance, in no less than three-way integration of mobile devices, local-ized, edge, and centralized cloud data centers. In this proposed ecosystem tasks to be offloaded are strictly encrypted and coded before their offloading and preserving sen-sitive user information as well as mathematically minimizing the encoded payload to the minimum. Also, we introduce a new hybrid approach that combines the data com-pression algorithms and game-theory-based optimization of tasks offloading, dynamic synchronization processes, and protocols of secure transmission through the encrypt-ed and encoded task packing. Finally, this study can be used in regards to the basic development of more sustainable, secure, and cost-efficient mobile-cloud systems that are highly essential in the next generation of Internet of Things (IoT) applications in cost-sensitive environments.

Abstract: Large Language Model (LLM)-based multi-agent systems have demonstrated strong capabilities in collaborative task-solving. However, a practical challenge emerges in extended collaboration: role drift, where agents gradually deviate from their designated responsibilities. This phenomenon manifests as boundary violations (e.g., a planner writing code), redundant work, conflicting decisions, and futile debates, ultimately degrading system performance. In this paper, we present RoleFix, a lightweight framework for detecting and repairing role drift in multi-agent collaboration. Our approach introduces: (1) a structured protocol requiring agents to declare their role, commitments, and dependencies at each turn; (2) a hybrid drift detector combining rule-based checks with LLM-based semantic judgment; and (3) a self-repair mechanism inspired by verbal reinforcement learning that triggers reflection, role reassignment, and execution resumption. Experiments on software engineering and research workflow tasks demonstrate that RoleFix reduces role drift incidents by 67.4% and improves task completion rates by 23.8 percentage points compared to baseline multi-agent systems, while introducing only 8.3% latency overhead.

Abstract: This study proposes a hybrid artificial intelligence framework for graduate subject allocation that enhances fairness, transparency, and operational efficiency in higher education. By integrating rule-based reasoning with advanced deep learning models, the framework supports data-driven decision-making for academic suitability, workload equity, and research alignment, while embedding Explainable Artificial Intelligence (XAI) to facilitate digital transformation in Thai universities. Traditional subject allocation processes in graduate programs are often manual, time-consuming, and subject to subjective judgment, thereby limiting their capacity to adapt to increasing curricular complexity and diverse faculty profiles. To address these challenges, the proposed framework combines institutional knowledge-based rules with machine learning techniques to model complex academic relationships across faculty, subjects, and workloads. Empirical analysis was conducted using data from 480 faculty members at Mahasarakham University and evaluated using multiple predictive models, including XGBoost, Wide-and-Deep Neural Networks, and Graph Neural Networks. Faculty performance was assessed using 20 institutional indicators reflecting teaching experience, research productivity, supervision, administrative responsibilities, and digital pedagogical competence. A multi-objective ranking algorithm was applied to simultaneously optimize academic suitability, workload balance, and research–teaching alignment. To enhance transparency and usability, the system incorporates a Faculty Subject Allocation Dashboard (FSAD) with SHAP-based interpretability, enabling administrators to understand and validate allocation decisions in real time. The results demonstrate that the proposed framework significantly improves allocation accuracy, workload equity, and decision transparency, offering a scalable and explainable solution for AI-driven academic governance.

Abstract: What are the numbers made of? More precisely, what are prime numbers made of? I posed this question to myself on the evening of August 19, 2025, which prompted prolonged introspection and profound contemplation. Then, I began constructing a numerical pyramid with prime numbers. The number one took the place of the central axis. Therefore, it is possible that large prime numbers could be surrounded by prime numbers on either side of one. However, this property extends to all even and odd non-prime numbers, but without one. The Goldbach ternary conjecture, which was proven by Harald Helfgott and is now recognized as the Goldbach-Helfgott theorem, is applicable to the observation that all odd non-prime numbers can be expressed as a sum of at least three prime numbers. This is due to the fact that non-prime numbers are a subset of all numbers greater than five. Once Goldbach's binary conjecture is proven, it will likely lead to the proof of Riemann's conjecture because we will be able to detect the structure of even numbers preceding prime numbers. For now, we can visualize this in the numerical structure of the first one trillion numbers and even further up to the largest known prime number. Let 3 203 431 780 337 be our number, which is verified as prime. If we subtract another prime number, 3 333 977 , from it, we obtain 3 203 428 446 360$. Subtracting one from the product verifies that 3 203 428 446 359 is prime. If so, then the sum of the two prime numbers plus one equals the proposed prime number above. This study has two objectives. First, it aims to present prime numbers as more than just their primality property. Second, it seeks to define the numbers 2 and 3 as a set of authentic prime numbers.

Abstract: Streaming Transformer Networks: Unified Hearing-to-Speech Recognition and Intelligent Text Generation Systems introduce a groundbreaking architecture that processes real-time audio streams to produce both synthesized speech outputs and contextually intelligent text, overcoming traditional limitations in multimodal AI systems. Traditional speech recognition models often operate offline, requiring full audio sequences before generating results, which hinders interactive applications. This work proposes a transformer-based framework that unifies hearing-to-speech translation directly converting input audio into natural-sounding speech with advanced text generation capabilities, enabling seamless dual-mode responses in conversational agents. By adapting transformers for streaming via causal attention and triggered mechanisms, the system achieves low-latency performance while maintaining high fidelity in prosody preservation and semantic coherence. Key innovations include shared encoder layers for efficiency, hybrid decoding paths for modality-specific outputs, and joint optimization across diverse objectives like word error rate minimization and perceptual quality enhancement. Evaluations on standard benchmarks demonstrate superior results, with latency under 200ms and error rates rivalling non-streaming baselines, paving the way for deployment in voice assistants, live captioning, and real-time dialogue systems. This unified approach not only reduces model complexity but also advances end-to-end learning for dynamic audio-to-multimodal generation tasks.

Abstract: This paper formulates and analyzes a novel compartmental model to study the spatial dynamics of corruption, framed as a pathogenic social strategy within a biological resource-competition framework. The model incorporates a renewable resource, whose scarcity drives the transmission of a corrupt strategy among a population of cooperators. The population is stratified into Cooperators (S), Corruptors (C), and Immunes/Enforcers (I), interacting within and between two connected patches via migration. The model exhibits a resource-dependent transmission rate and predator-prey dynamics between Corruptors and Enforcers. We establish the well-posedness of the coupled two-patch system by proving the positivity and boundedness of solutions. The system exhibits a corruption-free equilibrium, whose local and global stability is determined by patch-specific basic reproduction numbers R₀⁽¹⁾ and R₀⁽²⁾ , as well as a system-level reproduction number R0 that incorporates migration. We derive critical migration thresholds where the stability of the corruption-free state changes. Bifurcation analysis reveals the existence of a forward transcritical bifurcation at R₀ = 1, implying that reducing the system-level reproduction number below unity is sufficient to eliminate the corrupt strategy even in connected populations. Sensitivity analysis via Partial Rank Correlation Coefficients (PRCC) identifies the most critical parameters influencing R₀⁽ⁱ⁾ , providing evidence-based policy insights. Numerical simulations corroborate our analytical findings and explore the impact of asymmetric migration on the persistence of corruption. This work provides a theoretical foundation for understanding corruption through an ecological and spatial lens, highlighting the paramount importance of resource availability, enforcement mechanisms, and cross-border connectivity.

Abstract: This paper examines the strategic manipulationof government debt data—referred to as ”fiscal rule stretch-ing”—among European Union member states, particularly inresponse to electoral incentives, financial market stress, and theconstraints of the Stability and Growth Pact (SGP). Using Euro-stat revisions of government debt figures, we find that countrieswith higher debt levels, particularly those exceeding the 60% ofGDP SGP threshold, are more likely to have their debt figuresrevised upwards. Our analysis further reveals that eurozonemembers are more inclined to engage in rule stretching, especiallywhen facing Excessive Deficit Procedure (EDP) enforcement.Election timing plays a crucial role in this behavior, with un-scheduled elections and proximity to elections both significantlyincreasing the likelihood of debt revisions. Financial market stressamplifies these effects, as governments under pressure may resortto optimistic accounting to present more favorable debt statistics.While fiscal transparency appears to be positively associated withdebt revisions, suggesting that transparent governments may relyon more sophisticated methods of rule stretching, GDP growthdoes not show a significant impact. Overall, our findings highlightthe intersection of domestic political cycles, financial constraints,and EU-level oversight in shaping fiscal reporting practices acrossmember states.

Abstract: Hate speech on social media reproduces norms of inequality and gender stereotypes, disproportionately affecting women. This study proposes a hybrid approach that integrates emotional tone classification with explicit hostility detection to strengthen preventive moderation. We constructed a corpus from three open datasets (1,236,371 records; 1,003,991 after ETL) and represented the text using TF-IDF and contextual RoBERTa embeddings. We trained individual models (RoBERTa fine-tuned, Random Forest, and XGBoost) and a stacking metamodel (Gradient Boosting) that combines their probabilities. On the test set, the ensemble outperformed the base classifiers, achieving accuracy of 0.93 in hate detection and 0.90 in emotion classification, with an AUC of 0.98 for emotion classification. We implemented a RESTful API and a web client to validate the moderation flow before publication, along with an administration panel for auditing. Performance tests showed viability under moderate loads and concurrency limitations starting at 300 users, associated with deployment via an Ngrok tunnel. In general, the results indicate that incorporating emotional tone analysis improves the model's ability to identify implicit hostility and offers a practical way to promote safer digital environments. The probabilistic results obtained by the ensemble model were subsequently analyzed using the Bayesian Calibration and Optimal Design under Asymmetric Risk (BACON-AR) framework, which serves as a mathematical post-hoc validation layer to optimize the decision threshold under unequal costs. This framework does not modify the trained architecture but adjusts the estimated probabilities and selects the threshold that minimizes the total expected risk. By combining TF-IDF and RoBERTa embeddings with a stacked metamodel, the ensemble's decision function was optimized via regularization, improving generalizability and the stability of predictions. The incorporation of the BACON-AR framework strengthened the system's probabilistic consistency, ensuring that final decisions were aligned with the actual consequences of errors under an asymmetric risk scheme.

Abstract: This paper introduces a novel Transformer-Driven Pipeline that seamlessly integrates acoustic hearing, automated speech transcription, and writing synthesis into a unified end-to-end framework powered by advanced transformer architectures. Beginning with raw acoustic inputs captured via microphones, the pipeline preprocesses audio signals into spectrogram representations, leveraging stacked transformer encoders with multi-head self-attention to extract contextualized phonetic and prosodic features. These features feed into a sequence-to-sequence transcription module, where cross-attention mechanisms align auditory patterns with linguistic tokens, achieving robust speech-to-text conversion even in noisy environments or with diverse accents. Extending beyond transcription, the system employs a generative decoder to synthesize structured written outputs, such as summaries, reports, or formatted notes, by refining transcripts through autoregressive language modelling while preserving semantic fidelity and stylistic nuances derived from the original speech. Experimental validation on benchmark datasets like LibriSpeech and Common Voice demonstrates superior performance, with word error rates reduced by up to 25% compared to RNN baselines and enhanced fluency in synthesis metrics like BLEU scores. The pipeline's parallelizable design ensures real-time efficiency, making it ideal for applications in assistive technologies, live captioning, and automated documentation. This work highlights transformer's versatility in bridging auditory perception and textual production, paving the way for scalable multimodal AI systems.

Abstract: The application of quantum machine learning (QML) to security-relevant problems has attracted growing attention, yet its practical behavior in realistic workloads remains insufficiently characterized. This paper investigates the feasibility and limitations of variational quantum classifiers (VQCs) in a representative side-channel analysis (SCA) setting focused on distinguishing real from dummy power traces. A controlled benchmark framework is developed to examine training stability, sensitivity to key design parameters, and resource–performance trade-offs under realistic constraints. Hardware-relevant factors, including finite measurement budgets and device noise, are incorporated to approximate execution beyond idealized simulation and to quantify inference-time robustness. Experimental results indicate that VQCs can extract meaningful discriminative patterns from structured side-channel inputs, while robustness and performance depend on encoding choices, circuit depth, and measurement conditions. These findings provide an empirically grounded perspective on the applicability and limitations of QML in side-channel security and offer practical guidance for future research in this area.