Search Results (1,329)

This case study examines changes in participants’ educational perceptions and technology acceptance following a targeted, short-term training intervention based on the Technological Pedagogical Content Knowledge (TPACK) framework. Utilizing a quasi-experimental longitudinal pre-test/post-test design, the study examined the perceptions of 476 pre-service and in-service teachers at a major public university in Italy before and after a 10 h training course covering AI architecture, Large Language Models (LLMs), and practical pedagogical applications. Data were collected using a validated 4-point Likert-scale questionnaire, with relevant items subsequently mapped onto Technology Acceptance Model (TAM) constructs for the present analysis. Inferential analyses identified statistically significant positive shifts in perceptions of AI’s educational usefulness and practical application, although effect sizes varied across constructs. Increased variability in perceptions of automated tutoring suggested continuing divergence among participants rather than uniform acceptance. The findings suggest that structured technical and pedagogical AI literacy may support a shift from generalized concerns or uncertainty toward more informed and reflective engagement with AI, providing a useful foundation for the integration of AI literacy into teacher education frameworks. Full article

Industrial low-altitude multi-object tracking is challenged by overhead viewpoints, dense small targets, mutual occlusion, homogeneous appearances, and domain-shifted backgrounds. Vision-language pre-trained models provide useful semantic priors, but their category-level representations can mislead instance-level association when many same-class targets are densely distributed. This study develops an OC-SORT-based framework with reliability-aware semantic gating and retrieval-augmented trajectory memory. On the MOT17 reporting split (MOT17-09/10/11/13) with FRCNN detections, the FullModel achieves 46.41% MOTA, 40.00% IDF1, 34.542% HOTA, 26.686% AssA, and 382 identity switches for seed 42. Relative to fixed-weight CLIP+FAISS, it reduces identity switches and fragmentations while trading 0.84 MOTA points for a 0.68-point IDF1 gain; exact reruns with three seeds reproduce the same MOTA, IDF1, and IDSw totals. The VisDrone-derived subset is used only as an industrial low-altitude proxy stress test: its high precision coexists with very low recall, so the result exposes domain-shift limitations, rather than demonstrating deployment-level industrial performance. Full article

During sudden-onset disasters, social media text and images provide time-critical evidence for emergency awareness, damage assessment, and humanitarian response, but their cues are often incomplete, noisy, or conflicting. Direct equal-status fusion of heterogeneous pretrained representations can introduce semantic misalignment and allow unreliable evidence to influence the classifier. This paper proposes foundation-augmented reliability-conditioned dynamic adaptive fusion (FA–RC–DAF) for multimodal disaster classification. The model first preserves CLIP as a unit-weight text–image alignment base, maintaining a stable cross-modal decision space. It then projects BERTweet and SigLIP features into the CLIP-aligned space as bounded residual corrections, enabling domain-specific linguistic and complementary vision–language cues to refine the base without overwriting it. A confidence–agreement router estimates sample-level residual reliability from normalized-entropy predictive concentration and cross-encoder agreement, selectively admitting each residual before fusion. Explicit cross-modal interaction is performed only after the two streams have been reliability-refined. On CrisisMMD, FA–RC–DAF achieves 92.26% accuracy, 92.25% weighted F1, and 91.26% macro F1 under the retained five-class protocol. The protocol-aware published comparison is reported separately from the five-seed internally matched architectural comparison, which provides the primary evidence for method-level claims. Additional evaluations show differentiated behavior under event- and disaster-type shifts, stronger difficulty under forward temporal drift, and condition-dependent sensitivity to corrupted or unavailable inputs, providing a more fully characterized basis for multimodal disaster decision support. Full article

Semantic segmentation for autonomous driving is challenged by limited field-of-view (FoV), where occlusions and restricted viewpoints reduce available visual information. Existing approaches often incur considerable computational overhead, which is unsuitable for autonomous driving systems. To build robust and lightweight semantic segmentation models, attention-based knowledge distillation (KD) can be an efficient alternative. Nevertheless, KD remains primarily visual and provides limited information when visual cues are incomplete or ambiguous. In contrast, biological visual systems integrate visual cues with contextual and semantic knowledge to support robust scene perception. Motivated by this biomimetic principle, language-guided methods provide high-level semantic knowledge, offering another promising direction for enhancing visual understanding. However, the effectiveness of attention-based KD under limited FoV conditions and its interaction with semantic knowledge remain largely unexplored. To systematically investigate this problem, we devise Semantic-guided Attentive Feature Distillation (SAFD), a KD-based framework that combines attention-based feature distillation with semantic guidance derived from large language models (LLMs) without introducing additional inference-time computation. Through comprehensive analyses of multiple distillation strategies and textual representations using the framework, we observe that semantic guidance generally provides additional benefits when the gains from visual distillation are limited, while its effectiveness varies across different attention-based distillation strategies and textual representations. Furthermore, our analyses reveal that different attention mechanisms exploit semantic guidance in distinct ways, with consistent benefits across different pretrained text encoders and improved generalization. These findings provide practical insights into integrating linguistic semantics with visual knowledge transfer under limited FoV. Full article

OCR for low-resource languages is still held back by the same small number of issues: too little labeled image-text data, too few benchmarks, and thin language-specific tooling. Kurdish Sorani is a particularly awkward case. It is written in a modified Arabic script, runs right to left, and has orthographic habits that standard Arabic OCR engines handle poorly. This paper describes a transformer OCR system for Sorani trained almost entirely on synthetic data, meaning line images rendered on the fly from a text corpus rather than manually transcribed scans. The pipeline has three parts: corpus-driven line synthesis, a deterministic script-aware normalization step based on character-level transliteration, and a TrOCR encoder–decoder recognizer. Text lines are rendered with randomly sampled fonts and sizes, then passed through stochastic augmentation to mimic realistic distortions. The system is evaluated twice. On an in-distribution synthetic set of 200 rendered lines, the best model reaches a character error rate of 0.0434, a word error rate of 0.1246, and 64.0% exact matches. More importantly, on a real-world test set of 19 scanned Kurdish documents (467 lines, 28,468 characters) processed end-to-end through detection and recognition, it reaches a character error rate of 0.0305 and a word error rate of 0.1770, beating both Arabic and Kurdish Tesseract baselines and an existing Kurdish TrOCR model while being considerably smaller than the latter. A controlled ablation, in which eight variants are trained under one shared budget and scored on identical images, then isolates what each design choice contributes. The label space is the largest design effect, and the reason is concrete: the decoder’s pre-trained tokenizer has no representation for seven common Sorani graphemes, which cover 14.7% of the corpus and place a floor under any model trained on native-script labels. Corpus size dominates overall and behaves as a threshold, font diversity helps with diminishing returns, and stochastic augmentation buys robustness at a small cost in in-distribution accuracy. Aligning detected lines against the transcribed ones further shows that line detection contributes under 1% of the reported character error on this material. The broader point, at least for Sorani, is that the synthetic training data and the label space in which the model predicts have to be designed together: a compact recognizer built that way outperforms a substantially larger released Kurdish model on genuine document images. Full article

As the culture and tourism industry moves toward high-quality development, tourist perception has become a core dimension for measuring that quality. Most existing work on tourism spatial differentiation still looks at the supply side, and describes the pattern of tourism attractions. It pays limited attention to how supply and demand couple with each other, and it offers little systematic evidence from large-scale, highly heterogeneous regions. This study combines multi-source tourism point data with review texts from Trip.com (hereafter, Trip) and Rednote (Xiaohongshu). Using RoBERTa-wwm, a pretrained Chinese language model with whole-word masking, together with a coupling-coordination model and the geographical detector, we analyze how the spatial differentiation of tourism attractions couples with tourist perception, and we trace the driving mechanism to land resources, tourism resources and socioeconomic conditions. The results show that tourist sentiment perception in Xinjiang forms a spatial pattern that is led by core cities and declines outward layer by layer, and the coupling coordination between the distribution of tourism attractions and tourist sentiment perception remains higher in the north and lower in the south. The two systems do not, however, follow a simple linear relation: infrastructure coverage emerges as the strongest correlate of coordination, whereas regional economy and population size explain it only weakly. This study therefore extends the demand-side perspective in research on tourism spatial differentiation, and it offers a scientific basis and a decision-making reference for large-scale regions that seek to optimize tourism spatial layout and to upgrade the industrial quality. Full article

Accurate punctuation and capitalization are essential for the readability, interpretability, and structural coherence of machine-generated text. Their absence is particularly problematic in automatic speech recognition outputs and other forms of unstructured text, where missing punctuation and incorrect capitalization reduce both human readability and the effectiveness of downstream natural language processing tasks. This study proposes a hybrid XLM-RoBERTa–BiLSTM model for joint punctuation restoration and text capitalization on English-language data. The proposed architecture combines contextual representations produced by the multilingual pre-trained XLM-RoBERTa encoder with the sequential modeling capabilities of a bidirectional long short-term memory layer, followed by token-level classification in a unified label space. The model was trained and evaluated on a dataset derived from the IWSLT 2012 TED Talks corpus. Experimental evaluation on a held-out random test subset demonstrates strong performance. Excluding the dominant no-punctuation class, the model achieves an accuracy of 0.929, precision of 0.892, recall of 0.914, and an F1-score of 0.903. Including the dominant no-punctuation class increases these values to 0.961, 0.927, 0.919, and 0.923, respectively, reflecting the pronounced class imbalance in the dataset. Class-wise analysis shows high effectiveness for frequent punctuation classes and reliable capitalization prediction, whereas rare punctuation–capitalization categories remain more challenging because of their limited representation in the training and test subsets. Overall, the proposed hybrid XLM-RoBERTa–BiLSTM architecture achieves strong performance in joint punctuation restoration and text capitalization and represents an effective approach to improving the readability and structural quality of automatic speech recognition transcripts and other machine-generated text. Full article

Video anomaly detection has two very different answers to the question of how a detector should acquire its notion of “normal” for a given camera: adapt its parameters to hours of normal footage recorded by that exact camera, or perform no target-scene adaptation at all and rely on frozen, heavily pretrained backbones—a vision-language image–text model and a large language model applied to text descriptions of frames. We call the second setting target-scene-training-free: the backbones themselves are trained on very large general-purpose corpora, but no parameter is updated, fine-tuned, or transfer-learned on the target scene. That setting is far more convenient to deploy, but how much accuracy does it give up, and can any of that gap be closed without target-scene adaptation? We define an anomaly operationally, as a frame whose score under a scoring function $s(·)\in [0,1]$ exceeds a threshold, and evaluate every system with one metric definition and one implementation: frame-level area under the ROC curve (AUC) and equal error rate (EER), computed over all ground-truth-labeled test frames of each benchmark. We train a simplified future-frame-prediction network—a U-Net optimized with an intensity and gradient-difference loss, with the optical-flow and adversarial terms of the original design removed—on the normal-only training split of UCSD Ped1, UCSD Ped2, and CUHK Avenue, reaching a mean AUC of 0.848. A target-scene-training-free system starts far behind at 0.567, because its semantic signal is dropped in practice: the vision-language model that would produce it is too slow to run over a full test set. We close this gap in three ways, all evaluated on full test sets: CLIP-Guided Semantic Grounding (CSG), Vision-Language Reasoning (VLR), and Statistically-Calibrated Semantic Grounding (SCSG). Our best method raises mean AUC to 0.653, closing roughly a third of the gap with no target-scene adaptation. The trained model nonetheless leads on every benchmark, and because it is deliberately simplified, the gap we report is a conservative lower bound on the gap a fully engineered trained model would show. We report every result honestly, including a prompt-sensitivity analysis showing how much of the training-free result depends on prompt wording, cases where a more sophisticated method did not beat a simpler one, and a measured explanation of why UCSD Ped1 is the weakest benchmark for the trained model. Full article

Industrial equipment-maintenance knowledge is scattered across technical manuals, historical work orders, and field experience. Rapidly retrieving these heterogeneous sources and generating accurate guidance are essential to reducing downtime losses and ensuring operational safety. General-purpose large language models perform well in open-domain question answering, but their pretraining corpora lack industrial maintenance terminology and diagnostic reasoning patterns, causing severe mistakes in direct applications. Semantic-similarity-based retrieval-augmented generation can introduce domain knowledge to reduce hallucinations, yet it cannot preserve causal direction and procedural-order constraints. This limitation can logically misalign retrieved evidence and misguide maintenance decisions. Field queries also vary in wording and information completeness and often contain equipment abbreviations, colloquial transcriptions, and omitted key details, making fixed retrieval and response strategies unreliable. Our framework combines progressive fine-tuning with dual-path evidence fusion. A perplexity-driven sample-difficulty partition organizes two-stage fine-tuning so that the model learns basic domain knowledge before complex diagnostic reasoning, mitigating domain hallucinations. Parallel knowledge-graph and vector retrieval uses graph topology to preserve recorded relation directions and source procedures to supply operating-sequence information. A three-level router driven by retrieval necessity and entity-matching status selects direct answering, vector retrieval, or joint graph–vector retrieval for queries with different levels of information completeness. Independent tests, noisy-query experiments, and blinded expert evaluation show improved answer accuracy, logical consistency, and noise robustness. Relative to the base model, two-stage fine-tuning improves BLEU and ROUGE by approximately 14% and 7%, respectively. The results indicate that progressive domain learning and structured external evidence can provide reliable decision support for industrial equipment maintenance. Full article

Multimodal large language models (MLLMs) and vision–language models (VLMs) have rapidly entered medicine, demonstrating promising performance in clinical reasoning, radiology report generation, and visual question answering (VQA). However, many current multimodal architectures and pretrained visual backbones remain fundamentally rooted in two-dimensional (2D) image processing, even though major clinical imaging modalities, including computed tomography (CT), magnetic resonance imaging (MRI), optical coherence tomography (OCT), and echocardiography, are inherently volumetric or temporal. This narrative review examines the transition from 2D vision–language systems to volumetric multimodal AI, tracing the evolution from 2D and slice- or projection-based approaches through sequential and video-like methods to three-dimensional (3D) vision foundation models and native 3D VLMs/MLLMs. We examine their representational and computational trade-offs, evaluation gaps, and clinically grounded benchmarks. Approaches differ substantially in how they represent and preserve 3D information. Slice- and projection-based methods offer computational efficiency but may discard spatial context, whereas sequential and native volumetric approaches increasingly model relationships across the full imaging study. Recent 3D foundation models and multimodal systems demonstrate the feasibility of reusable volumetric representations and language-enabled 3D image interpretation, but face barriers in computational cost, training-data scale, evaluation methodology, and clinical reliability. Only 53% of Med-Gemini-3D reports were judged clinically acceptable, and natural language processing (NLP) metrics such as BLEU and ROUGE correlate poorly with diagnostic correctness. True 3D multimodal medical intelligence remains in its early stages. Future progress requires efficient volumetric representation strategies, clinically grounded evaluation frameworks, standardized benchmarks, and robust cross-institution validation. Full article

As the crisis of antibiotic resistance escalates, phage therapy has regained attention as an alternative strategy. Artificial intelligence (AI) technologies offer new avenues to overcome the bottlenecks inherent in traditional bacteriophage research. This review summarizes the multi-dimensional innovative applications of machine learning, deep learning, and large biological models in phage studies. In the fields of phage recognition and genomics, support vector machines (SVMs), convolutional neural networks (CNNs), and pre-trained protein language models can all achieve recognition accuracy rates of over 90%. Furthermore, tools such as DeepHost and VirSorter2 can efficiently identify phage sequences, annotate functional genes, and predict hosts at the species or strain levels. For clinical translation, AI integrates patient characteristics, bacterial phenotypes, and phage profiles to customize cocktail regimens for individualized phage therapy. Graph neural network-based models like DeepPBI-KG integrate multi-omics knowledge graphs to precisely predict phage-host interactions (PHIs), whereas agent-based simulation and defense protein predictors forecast phage resistance evolution. Additionally, generative AI can support the de novo design of functional phage genomes and mine massive unannotated virome dark matter. Nevertheless, this cross-disciplinary field faces significant constraints, including uneven and biased sequencing datasets, insufficient model interpretability, and dual-use biosafety ethical risks accompanied by unclear algorithm accountability and incomplete global supervision systems. Future research should optimize standardized multimodal databases, develop explainable AI algorithms, and establish cross-disciplinary ethical governance frameworks to facilitate closed-loop verification between computational prediction and wet-lab experiments. In conclusion, the deep integration of AI and phage biology provides revolutionary strategies to tackle multidrug-resistant infections and advances the clinical transformation of phage precision medicine. Full article

Automated alert triage can reduce Security Operations Center (SOC) workload, yet the validation-tuned thresholds deployed systems rely on carry no finite-sample control of their operational error rates and degrade unpredictably under distribution shift. We present a model-agnostic dual-threshold conformal deferral architecture: high-score alerts are auto-escalated under finite-sample marginal class-conditional control of the benign-escalation probability (budget α), low-score alerts are auto-closed under matching control of the threat-miss probability (budget β), and the rest are deferred to an analyst. It needs no retraining and closes an automatic zone rather than certifying what the calibration data cannot support. We evaluate it on a reinforcement-learning investigation agent in a simulated SOC and on four classifiers trained on CIC-IDS2017 and tested on CSE-CIC-IDS2018, using stratified 25,000-flow calibration and evaluation samples, with attack-type recall computed over the full 16.2-million-flow corpus. Pooling episodes from ten trained policies across two evaluation datasets, the architecture automated 73.7% of decisions at α = β = 0.01—a figure for that predefined pooled mixture rather than a per-policy or per-dataset guarantee—realizing benign auto-escalation and threat auto-close rates of 0.0099 and 0.0101 and deferring the hardest ~26% of alerts. After recalibration on labeled target-domain data, severe cross-dataset degradation appears not as a silent error but as sharply reduced certifiable automation, with deferral rising to 79–99% for the most affected classifiers. This visibility is a property of the recalibrated layer: thresholds left un-recalibrated after a shift continue to certify nothing while still deciding, so the architecture requires periodic recalibration on labelled target-domain alerts to deliver it. Substituting open-weight language models for the analyst inside the band failed a pre-specified criterion at every scale tested from 7B to 32B across two model families, with the discriminative signal flat in model size and far below the first-stage policy’s own. Full article

Part-of-speech (POS) tagging in Arabic is hard because its rich root-and-pattern morphology and the absence of short vowels make one unvoweled string compatible with many categories. This paper presents Jawhar, a hybrid framework that couples a high-performance morphological analyser with contextual reranking using a pretrained Arabic language model. Jawhar is an autonomous engine inspired by Al-Khalil MorphoSys and rebuilt in Python that replaces the original XML databases with optimised JSON structures for faster inference. It enumerates the morphologically valid candidates of each token, and a CAMeL-BERT stage then scores each candidate by its full morphological signature (type, POS, root, pattern, and voweled form). On the Prague Arabic Dependency Treebank, mapped to the universal 17-tag POS scheme, the fine-tuned scorer reached 96.4% token accuracy (macro-F1 0.921), on par with published neural taggers, while a candidate-constrained hybrid attached a full morphological analysis to 65.9% of tokens at the same accuracy and reached a 97.4% oracle ceiling. A rule-based configuration reached 54.4%, and the zero-shot reranker reached parity (54.3%), which showed that within a fixed candidate set, reordering could not cross the coverage ceiling. The main contribution was a token-level decomposition of the error budget that isolated candidate coverage and label mapping from contextual ranking, released with a public analyser and harness. Full article

Small, heterogeneous clinical datasets pose a challenge for whole-cohort prediction because clinically meaningful treatment or response patterns may be diluted across biologically diverse patients. We describe and evaluate NetraAI, an interpretable dynamical-systems framework for selective subgroup discovery that uses finite-iteration contraction-inspired dynamics and long-range memory (LRM) to identify stable, outcome-linked Model-Derived Subgroups (MDS). This system can abstain by assigning No Call when a stable subgroup assignment is not supported. A large language model (LLM) Strategist is outlined only as a possible future extension; it is not evaluated here and contributes nothing to the results reported. Foundation and language models asked to perform subgroup discovery directly did not recover the structure the specialized discovery step recovered. We demonstrate this framework across three retrospective clinical trial datasets: Clinical Antipsychotic Trials of Intervention Effectiveness (CATIE) schizophrenia (olanzapine vs. perphenazine comparative treatment-preference benchmark), Canadian Biomarker Integration Network in Depression (CAN-BIND) depression (escitalopram response), and Comprehensive Molecular Characterization of Advanced Pancreatic Ductal Adenocarcinoma for Better Treatment Selection (COMPASS) pancreatic cancer (GnP vs. FOLFIRINOX observational regimen-associated response). The benchmark is not a contest between NetraAI and competing predictors: the same eight downstream methods are evaluated with and without what NetraAI discovered. Given the full feature sets and their own selection procedures, those methods were at or near chance on all three datasets, and blind de novo searches by an independent interaction model and by a pretrained tabular foundation model did not recover an equivalent signature or subpopulation. In internal downstream evaluation, given the discovered variables alone—the same patients, the same classifiers, the full cohort and no abstention of any kind—every one of the eight methods improved on every dataset, 24 of 24 method-dataset comparisons, moving from a raw-feature range of 0.46–0.62 AUC to 0.54–0.78. Restricting further to the subpopulation in which those variables hold improved all eight methods again in CAN-BIND and in COMPASS, 16 of 16 comparisons, reaching 0.66–0.83 and 0.96–1.00, respectively; in CATIE, where the called subgroups are the least outcome-homogeneous of the three, it improved only one of eight. Taking the framework as a whole, 23 of 24 method-dataset combinations improved over the raw-feature baseline. NetraAI abstains on patients without stable subgroup structure, calling 27.7% to 40.4% of each cohort. The contribution demonstrated is therefore subgroup discovery rather than downstream prediction, and its beneficiaries are the conventional methods themselves. In COMPASS, NetraAI identified a three-SNV signature associated with regimen-linked response ranking among called patients; because the cohort was observational and the permutation test was conditional on the selected signature, this finding is exploratory. The two mechanisms are complementary rather than competing: variable discovery establishes which features carry the structure, and abstention establishes in which patients it holds. Neither mechanism replaces conventional modeling. Variable discovery improved every method on every dataset, the pretrained tabular foundation model included; identifying the population in which those variables hold conferred further benefit in two of the three datasets and not in the third. These findings support NetraAI as an exploratory system for generating compact, inspectable subgroup hypotheses that may inform future enrichment strategies after external validation, and indicate that its value lies in what it contributes to other methods rather than in competing with them. Full article

Bug localization is the task of automatically identifying the source files responsible for a reported defect. It is a critical step in software maintenance that accelerates defect resolution. Information retrieval (IR) methods are simple and effective at exploiting historical signals such as bug-fixing recency and frequency, but they struggle to bridge the lexical gap between natural-language bug reports and programming-language identifiers. Recent work increasingly leverages pre-trained language models (PLMs) for code to close this gap. However, current PLM-based approaches still rely on a single code encoder that ignores program structure and aggregates function-level signals into file-level representations via uniform pooling. We propose a dual-encoder bug localization (DuaLoc) framework that jointly addresses these limitations. DuaLoc combines two pre-trained language models: UniXcoder for the semantic understanding of source code and GraphCodeBERT for awareness of data-flow structure. Both encoders are fine-tuned with a contrastive objective that shapes the embedding space around the localization task. A bug-report-conditioned attention mechanism then aggregates function embeddings into query-dependent file representations. The resulting neural similarity scores are then fused with classical IR features in a learning-to-rank model. DuaLoc outperforms representative classical and PLM-based baselines across most evaluation settings on a widely used benchmark of six open-source Java projects. Full article