Articles in this Volume

Research Article Open Access
Cost-Aware Adaptive Prompt Selection for Large Language Models Using UCB and Thompson Sampling
Article thumbnail
This paper investigates cost-aware adaptive prompt selection for large language models by modeling four prompt strategies as arms in a multi-armed bandit. Standard UCB and Thompson Sampling are compared with contextual extensions, LinUCB and Linear Thompson Sampling (LinTS), on GSM8K. The study evaluates accuracy, cumulative reward, cumulative regret, average token usage, tokens per correct answer, and budget satisfaction. On the held-out evaluation set, the Global Best prompt achieved 96.8% accuracy with a cumulative regret of 3.00, while the Oracle upper bound reached 97.4%. LinUCB achieved 95.47% accuracy with an average of 200.18 tokens per question, reducing token use by approximately 9.9% relative to the Global Best prompt (222.14 tokens). Under the medium normalized token budget B = 0.2893, cost-constrained LinUCB maintained 95.47%-95.49% accuracy and 93% budget satisfaction, compared with 30%-35% for LinTS. In addition, 92% of evaluation questions produced identical binary rewards across all four prompts, indicating limited prompt-specific heterogeneity. The results suggest that contextual prompt selection may provide greater value for resource control than for accuracy improvement in homogeneous, high-accuracy reasoning settings.
Show more
Read Article PDF
Cite
Research Article Open Access
Short-Term Traffic Flow Prediction for Urban Roads Based on Machine Learning Algorithms
Article thumbnail
As cities grow and vehicle fleets expand, congestion on urban arterials has become harder to manage. Short-term traffic flow prediction sits at the center of many intelligent transportation applications, from adaptive signal control to route guidance. Observed flow is driven by clock time, day type, weather, road class, and speed profiles, and the resulting series shows strong nonlinearity together with daily and weekly cycles. Conventional statistical predictors often miss these patterns, while deep models can fit them but remain sensitive to learning rate, hidden width, and regularization. Manual grid search is slow and frequently stops at suboptimal settings. We therefore combine BiLSTM, Transformer, and differential evolution (DE) into one pipeline called DE-Transformer-BiLSTM: BiLSTM reads the sequence in both directions, Transformer models longer interactions, and DE searches hyperparameters without hand tuning. On a roadside dataset of 1,728 samples, the proposed model beats eleven reference algorithms on MSE, RMSE, MAE, MAPE, and R², which supports its use in operational traffic management.
Show more
Read Article PDF
Cite
Research Article Open Access
Optimization of Transparency and Accountability Structures in Algorithm-Driven Public Governance
Article thumbnail
Governance by algorithm is widely applied in the public sector for the purposes of distributing welfare, estimating administrative risks, regulating urban behavior, rationing public services, and enabling automated decisions. However, transparency and accountability mechanisms are often scattered in the technical, legal, and organizational dimensions. In this work, we present an empirical-computational approach to improving transparency and accountability designs within algorithmic governance architectures in the public sector. We construct a governance assessment dataset based on policy guidance, procurement requirements, data inventories, model cards, decision logs, explanation logs, appeal decisions, audits, and stakeholder review annotations. Our methodology combines natural language processing, rule-based compliance coding, accountability graph construction, supervised classification, risk scoring, and intervention simulation. Using PostgreSQL, Elasticsearch, Doccano, Neo4j, NetworkX, XGBoost, LightGBM, Optuna, and PyGAD, we analyze measures of disclosure completeness, explanation quality, procedure traceability, responsibility path consistency, appeal availability, correction timeliness, and audit resolution efficiency. We show that structural interventions increase the transparency completeness score from 0.642 ± 0.071 to 0.813 ± 0.058 (p < .001) and decrease the aggregate Transparency- Accountability Risk Score by −0.184 ± 0.039 (p < .001).
Show more
Read Article PDF
Cite
Research Article Open Access
A pplication and T echnical F ramework of L arge L anguage M odel in S ports F itness S cene
Large language models deliver strong performance in understanding natural language, generating knowledge content and sustaining multi-turn conversations, opening new technical possibilities for digital transformation within sports and fitness industries. This paper first sorts out two core technical components of such models: pre-training and fine-tuning. Drawing on existing published studies and real-world product cases, it investigates several typical usage scenarios, including fitness knowledge Q&A, custom exercise plan generation, AI-based virtual coaching, and psychological intervention for exercisers. Apart from summarizing the practical strengths brought by large-language-model-driven fitness services, this review also points out multiple tangible obstacles in real-world application. These include model-induced knowledge hallucination, insufficient high-quality datasets built for vertical fitness domains, unsatisfactory multimodal integration effects, as well as unsettled issues around user privacy protection and liability division. Corresponding countermeasures are discussed from five practical dimensions: domain dataset construction, targeted model optimization, multimodal technical upgrading, privacy risk control and industrial standard-setting. It should be noted that most current research remains at the exploratory phase, and real-user-based empirical evidence is still limited. This work hopes to offer usable references for follow-up academic research and actual product deployment of large language models applied in sports-fitness vertical domains.
Show more
Read Article PDF
Cite
Research Article Open Access
Research on Energy-Efficient Adaptive Processing Control Method for Industrial Robots Driven by Multi-Source Perception in Complex Working Conditions Toward Green Manufacturing
Article thumbnail
This paper provides an energy-efficient adaptive processing control algorithm design for robotic systems functioning in intricate grinding conditions. The testing system includes the following elements: ABB IRB 2600 industrial robot, ATI Mini45 six-axis force sensor, Basler acA1920-40gm industrial camera, Yokogawa WT310 power meter, and National Instruments USB-6363 data acquisition unit. Multiple signals – force, vision, electrical, and robotic movement – are acquired simultaneously in equal intervals of time. The robot performs 180-second grinding operations of 36 stainless steel test specimens possessing smooth, bent, and rough surfaces. Filtering is performed for the force and power signals (with frequency of 10 Hz) while images are processed with rate of 20 Hz. Logs of robotic movement are aligned by means of 0.1 seconds. Adaptive adjustment of feed rate, reference contact force, and rotation speed of the tool is implemented by using MATLAB/Simulink and RobotStudio tools. The adaptive algorithm results in reducing energy consumption by 13.8%, on average, compared to the case of constant control parameters (standard deviation of 2.6%, -13.8 ± 2.6%) while peak power decreases by 118.4 W (standard deviation of 24.7 W, -118.4 ± 24.7 W).
Show more
Read Article PDF
Cite
Research Article Open Access
Deep Reinforcement Learning in Communication Networks
In communication networks, during different traffic and channel conditions, scarce resources such as spectrum, power, bandwidth and computing resources need to be allocated. In this review, the authors examine how Deep Reinforcement Learning (DRL) can support such decisions, what its role is and why there is a need for it. The representative tasks and decisions made in the studies are discrete wireless allocation, continuous congestion control, hybrid edge offloading, graph-based routing, adaptive video streaming, and network slicing. Observations, actions, objectives, algorithm design, strengths and limitations are explored and contrasted between the tasks in each application. From the literature, one can see that Deep Q-Network agents can handle local wireless decision making; however, fine action choices and the multi-agent interactions lead to a higher coordination cost. Actor–critic controllers adjust sending rates according to feedback, but there are issues of fairness and transferability to unknown traffic. Although graph-based routing does not guarantee robustness to changes in demand, hybrid offloading reduces the search burden while retaining optimization overhead. DRL is most compelling when conditions change and adaptation is required, and when benefits can be shown based on appropriate networking baselines. Future research topics include data coverage, enforceable service constraints, as well as coordination among controllers of different time scales.
Show more
Read Article PDF
Cite
Research Article Open Access
Paradigm Evolution and Engineering Selection of Retrieval-Augmented Generation for Large Language Models
Article thumbnail
Large language models (LLMs) suffer from two persistent problems in their parametric knowledge: update lag and factual deviation, which hinder their deployment in professional domains. Retrieval-augmented generation (RAG) connects external corpora at inference time, expanding the knowledge boundary without modifying model parameters, and has become the mainstream solution for enterprise knowledge bases, intelligent question answering, and document understanding. This paper divides RAG technology into four stages along its design focus: naive retrieval, precise retrieval, adaptive routing, and agentic reasoning, and examines the principal contradictions and new costs of each stage. On this basis, a hierarchical evolution framework is proposed that cross-compares six levels—retrieval, knowledge, fusion, generation, conflict arbitration, and engineering evaluation—against the four stages, identifying gaps in existing research, especially the unresolved arbitration of multi-source contradictory documents across all stages. The paper also discusses applicable boundaries: when knowledge is already well mastered by the model, when the context window suffices to cover the full corpus, or when the knowledge base is small and sparsely updated, adding a retrieval link only increases latency and noise. Finally, selection guidance is provided along three dimensions: accuracy, computational cost, and robustness, serving as a reference for engineering design and future research.
Show more
Read Article PDF
Cite
Research Article Open Access
Multi-Source Data Fusion Based Market Demand Forecasting and Intelligent Business Decision Optimization
Article thumbnail
New retail demand is determined by unpredictable sales dynamics, inventory level, promotion actions, weather conditions, and time factors, whereas stockouts may lead to demand underestimation due to distorted sales and negative replenishment decision-making. In order to solve the issue, a multi-source forecasting and business decision system has been designed with the help of FreshRetailNet-50K. First, stockout-affected demand was restored, followed by encoding and dynamic fusion of historical demand, inventory, promotion, weather, calendar, and product data via gated attention network to obtain seven days point and quantile forecasting. After that, forecasting was integrated in the rolling replenishment strategy with regard to ordering, holding, spoilage, stockout, capacity, shelf-life, and lead-time costs. According to experimental results, the proposed multi-source model yielded an MAE score of 6.84, an RMSE score of 10.27, and WAPE value of 29.36% along with the cost index reduction of the replenishment strategy from 424 to 326. The stockout ratio was lowered to 5.8%, spoilage ratio – to 3.7%, inventory turnover became equal to 8.6, and demand fulfillment to 94.2%. Consequently, one may state that incorporation of heterogeneous market data into probabilistic forecasting may lead to better demand estimation and operational efficiency. Thus, the system offers practical grounds for intelligent retail decision optimization, maintaining the link between forecast accuracy and inventory management metrics.
Show more
Read Article PDF
Cite
Research Article Open Access
When Language Rewrites History: Quantifying and Mitigating Cross-Lingual Inconsistency in LLM-Based Historical Question Answering via Quantitative Analysis and LLM Fine-Tuning
Article thumbnail
Open-ended historical questions can elicit different responses from a large language model when only the input language changes. We tested this with 200 questions about the Great Leap Forward and related modern Chinese history, written in English, Chinese, and Japanese. ChatGPT and Gemini generated 1,200 answers. We compared each answer with a reference in the same language using Emotion Risk (ER) and Embedding Distance (ED), combined these measures in a Final Score, and measured score variation across languages. Both models had their highest mean Final Scores for Japanese questions: 39.04 for ChatGPT and 39.68 for Gemini. We then split the questions by ID and fine-tuned Qwen2.5-1.5B-Instruct with supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). On 90 held-out language-level samples, mean Final Score fell from 34.55 in the original model to 34.36 after SFT and 31.88 after DPO. SFT gave the lowest mean Cross-Language Disparity (5.12, compared with 6.90 originally and 5.91 after DPO). These results show measurable language-conditioned differences in this historical question set and a reduction in the proposed response-risk score after training. The automatic metrics do not establish factual accuracy or historical bias on their own.
Show more
Read Article PDF
Cite
Research Article Open Access
DSA-YOLO for Foreign Object and Surface Defect Detection on Mining Conveyor Belts
Article thumbnail
Small foreign objects and elongated defects are difficult to detect in mining conveyor belt images. DSA-YOLO extends YOLOv8n through directional gated residual features, a stride-four branch with prediction towers shared across four scales, and bounded area-balanced regression. Experiments use 2,148 images with 14,700 boxes, acquisition-group isolation, and three random seeds. DSA-YOLO achieves 83.10% mAP50 and 54.63% ± 0.33% mAP50:95, improving on YOLOv8n by 9.47 and 6.36 percentage points. Foreign-object recall reaches 80.06% ± 0.87%. With 3.18 million parameters, mean end-to-end latency is 2.78 ms on an RTX 4090 at FP32 and batch size one, an increase of 0.32 ms. The results support joint detection under the evaluated conditions; transfer to independent sites remains to be assessed.
Show more
Read Article PDF
Cite