Search Results (146)

Public deployment of generative AI in cultural institutions is constrained by small, rights-restricted datasets, strict latency and runtime-stability requirements, and limits on visitor-data collection. This study presents a deployment-oriented framework for adapting a pre-trained text-to-image diffusion foundation model to a heritage-specific visual domain using Low-Rank Adaptation (LoRA). A Stable Diffusion v1.5 backbone is specialized through data-centric curation and LoRA fine-tuning, then served through an asynchronous edge architecture that links a Unity client and a local Python (version 3.10) inference server for public-facing operation on a native 400 × 1080 vertical canvas. To support deployment decisions without collecting personally identifiable information, the system records only anonymous operational logs and evaluates sustained-load behavior under repeated inference. In a 1000-iteration profiling test, the proposed configuration maintained stable runtime behavior without observable upward memory drift, with a peak allocated VRAM of 3.04 GB and an average end-to-end latency of 3.12 s. An 8 h field deployment further indicated service continuity under public interaction, while a CLIP-based proxy analysis under matched prompts and seeds suggested improved relative style controllability after adaptation (0.848 vs. 0.799). Rather than claiming cultural authenticity or visitor-level effects, this study offers a data-centric, deployment-oriented methodology for operating public-facing generative AI under small-data, latency, and privacy constraints. Full article

Breast cancer histopathology classification remains a fundamental challenge in computational pathology due to variations in tissue morphology across magnification levels. Convolutional neural networks (CNNs) have long been the standard for image-based diagnosis, yet recent advances in vision-language models (VLMs) suggest they may provide strong and transferable representations for complex medical images. In this study, we present a systematic comparison between CNN baselines and large VLMs—Qwen2 and SmolVLM—fine-tuned with Low-Rank Adaptation (LoRA; $r=16$, $\alpha =32$, dropout = $0.05$) on the BreakHis dataset. Models were evaluated at $40\times$, $100\times$, $200\times$, and $400\times$ magnifications using accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). While Qwen2 achieved moderate performance across magnifications (e.g., 0.8736 accuracy and 0.9552 AUC at $200\times$), SmolVLM consistently outperformed Qwen2 and substantially reduced the gap with CNN baselines, reaching up to 0.9453 accuracy and 0.9572 F1-score at $200\times$—approaching the performance of AlexNet (0.9543 accuracy) at the same magnification. CNN baselines, particularly ResNet34, remained the strongest models overall, achieving the highest performance across all magnifications (e.g., 0.9879 accuracy and 0.9984 AUC at $40\times$). These findings demonstrate that LoRA fine-tuned VLMs, despite requiring gradient accumulation and memory-efficient optimizers and operating with a significantly smaller number of trainable parameters, can achieve competitive performance relative to traditional CNNs. However, CNN-based architectures still provide the highest accuracy and robustness for histopathology classification. Our results highlight the potential of VLMs as parameter-efficient alternatives for digital pathology tasks, particularly in resource-constrained settings. Full article

As a reliable and high-potential wireless communication technology for the Internet of Things (IoT), LoRa excels in long-distance and low-power transmission. The star topology adopted by traditional LoRaWAN suffers from poor deployment flexibility and insufficient scalability in scenarios with complex terrain or harsh environments. LoRa-Mesh networks can effectively solve coverage challenges through characteristics such as multi-hop and self-organization; however, the relay and forwarding requirements of nodes also introduce new challenges in energy consumption management. To address the energy consumption management challenges of LoRa-Mesh, this paper proposes a Data-Driven Energy Saving (DDES) protocol. It flexibly sets and dynamically fine-tunes node sleep durations based on data changes, constructs an efficient energy-saving framework through uplink data streams, and implements precise control over nodes via downlink post-analysis messages to achieve on-demand energy saving. Simulation results in the smart agriculture scenario of soil moisture monitoring and irrigation show that compared with protocols without a sleep mechanism, the battery life of the LoRa-Mesh network using the DDES protocol is extended by approximately 20 times. The proposed protocol breaks through the limitations of fixed sleep schemes, realizes refined and flexible division of sleep regions, and exhibits significant advantages in LoRa network energy saving. Full article

Accurate species identification of necrophagous flies is fundamental to forensic entomology, particularly for postmortem interval (PMI) estimation in decomposed remains. Here, we conducted a targeted carrion-baited survey along the Shandong Peninsula and documented 15 Sarcophaga species, including the first regional records of S. cinerea, S. pingi, and S. pterygota. We established an expert-validated image dataset for automated identification. We then developed a parameter-efficient identification framework by fine-tuning a pretrained Vision Transformer with Low-Rank Adaptation (ViT-LoRA) on this custom dataset. Compared with conventional CNN-based models, ViT-LoRA achieved 98.50% species-level accuracy while updating only ~0.16 M trainable parameters, and it converged rapidly and stably within ~10 epochs, demonstrating efficient adaptation under limited training data. This study provides faunistic and distributional data on carrion-associated Sarcophaga species in the coastal Shandong Peninsula, characterizes their regional distribution patterns, and offers a scalable image-based identification approach for forensically important sarcophagid flies. Full article

Accurate rooftop segmentation from aerial imagery is essential for large-scale urban analysis, including applications such as solar potential assessment and urban monitoring. However, it remains constrained by the high cost of dense annotation and the limited generalisation of supervised models across heterogeneous urban morphologies. This study investigates binary rooftop segmentation for fine-tuning large image-editing foundation models using parameter-efficient Low-Rank Adaptation (LoRA). Using parts of Brisbane metropolitan dataset (split 80/20 into 97 training and 24 testing tiles), three paradigms were evaluated under a unified protocol: zero-shot image-editing models (including Gemini 3 Pro), a segmentation-first baseline (Segment Anything Model 3, SAM3), and LoRA-adapted diffusion models (FLUX.1 Kontext, FLUX.2, and Qwen Image Edit 2509) fine-tuned each 250 steps up to 5000 steps. Evaluated under zero-shot conditions, the generative models demonstrated varying levels of boundary fidelity. The Gemini model achieved a strong zero-shot baseline with [IoU, Dice] scores of [85%, 91%], followed by the SAM3 baseline, which also achieved a stable [84%, 91%] but exhibited increased false negatives in visually complex scenes. The tested diffusion models (FLUX.1 Kontext, FLUX.2, and Qwen) showed more limited initial spatial overlap, scoring [45%, 55%], [67%, 78%], and [33%, 46%], respectively. Following LoRA adaptation, the FLUX and Qwen models showed substantial improvements, with their respective [IoU, Dice] metrics increasing to [89%, 94%], [82%, 90%], and [87%, 93%]. FLUX.1 Kontext achieved the strongest overall performance at step 4250, yielding a mean IoU of 89% (SD = 3.16%) and a pixel accuracy exceeding 96%. These results demonstrate that parameter-efficient fine-tuning, combined with rigorous evaluation under class-imbalanced conditions, can transform general-purpose generative models into competitive, scalable spatial analysis tools that match or exceed both dedicated segmentation baselines and strong zero-shot multimodal models. Full article

Large language models (LLMs) have shown strong performance in open-domain dialogue, yet they continue to struggle with long-term multi-session conversations (MSC), particularly in non-English languages such as Korean. In this work, we present a comprehensive empirical study on enhancing Korean MSC capabilities of LLMs through dataset construction, memory modeling, and parameter-efficient fine-tuning. We introduce an extended Korean MSC dataset that explicitly distinguishes between persona memory (long-term user attributes) and episode memory (short-term, event-driven information), enabling more effective memory management across sessions. Using this dataset, we evaluate LLM performance on three core MSC tasks: session summarization, memory update, and response generation. Our experiments reveal that Korean MSC is intrinsically more challenging than English MSC and that memory update and response generation require substantial reasoning ability. To address these challenges, we compare LoRA, DPO, MoE, CPT, Layer Tuning, and neuron-level tuning methods. Results consistently show that neuron tuning, guided by a novel language-specific neuron identification method based on activation scores and entropy, achieves superior performance and robustness, particularly in continual learning settings. Overall, our findings highlight neuron-level adaptation as an effective and interpretable approach for improving long-term conversational ability in low-resource languages. Full article

To meet the growing demand for autonomous decision-making and real-time optimization in industrial manufacturing, integrating Artificial Intelligence-Generated Content (AIGC) services with Industry 5.0 can enable real-time industrial intelligence. The effectiveness of a generative model is closely related to the current state of the production environment. However, existing studies often ignore the dynamic temporal relationship between generative models and production environments, especially in industrial scenarios with large model transmission delays and random AIGC task arrivals. Therefore, we define a novel metric, namely the Age of Model (AoM), to measure the freshness of generative models with respect to current industrial tasks. We then formulate an average-AoM-minimization problem that jointly considers LoRA-based fine-tuning, wireless transmission and resource allocation. To solve this problem, we propose a Hybrid-Action Multi-Agent Proximal Policy Optimization (HA-MAPPO) algorithm. The proposed algorithm follows the centralized training and decentralized execution (CTDE) paradigm and introduces a Main-Agent Priority State Strategy to support coordinated training and independent execution. In addition, a multi-head output structure is designed to handle the hybrid-action space, which includes discrete fine-tuning association decisions and continuous transmission resource allocation. Simulation results show that the proposed scheme outperforms all benchmark methods. Specifically, the cumulative rewards are improved by approximately $11.13\%$, $20.32\%$, $36.61\%$, and $38.78\%$ compared with the four benchmark algorithms, respectively. These results demonstrate that the proposed scheme can significantly reduce the average AoM while providing high-quality and timely industrial AIGC services. Full article

The proliferation of fake reviews on e-commerce and social platforms has severely undermined consumer trust and market integrity, necessitating robust and interpretable real-time detection mechanisms with multi-sensor data fusion capabilities. While traditional machine learning approaches have shown promise in identifying fraudulent reviews, they often lack transparency and fail to leverage the rich contextual knowledge embedded in large-scale datasets. In this paper, we propose KE-MLLM (Knowledge-Enhanced Multimodal Large Language Model), a unified framework that integrates knowledge-enhanced prompting with parameter-efficient fine-tuning for explainable fake review detection. Our approach employs LoRA (Low-Rank Adaptation) to fine-tune lightweight large language models (LLaMA-3-8B) on review text, while incorporating multimodal behavioral sensor signals including temporal patterns, user metadata, and social network characteristics for comprehensive anomaly sensing. To address the critical need for interpretability in fraud detection systems, we implement a Chain-of-Thought (CoT) reasoning module that generates human-understandable explanations for classification decisions, highlighting linguistic anomalies, sentiment inconsistencies, and behavioral red flags. We enhance the model’s discriminative capability through a knowledge distillation strategy that transfers domain-specific expertise from larger teacher models while maintaining computational efficiency suitable for edge sensing devices. Extensive experiments on two benchmark datasets—YelpChi and Amazon Reviews from the DGL Fraud Dataset—show that KE-MLLM achieves strong performance, reaching an F1-score of 94.3% and an AUC-ROC of 96.7% on YelpChi and outperforming the strongest baseline in our comparison by 5.8 and 4.2 percentage points, respectively. Furthermore, human evaluation indicates that the generated explanations achieve 89.5% consistency with expert annotations, suggesting that the framework can improve the interpretability and practical usefulness of automated fraud detection systems. The proposed framework provides a useful step toward more accurate and interpretable fake review detection and offers a practical reference for building more transparent and accountable AI systems in high-stakes applications. Full article

Low-Rank Adaptation (LoRA) has become a widely adopted parameter-efficient fine-tuning (PEFT) technique for large language models (LLMs). LoRA’s benefits stem from its light weight and modular adapters. Standard LoRA applies adapters uniformly across all Transformer layers, implicitly assuming that each layer contributes equally to task adaptation. However, LLMs are found to have internal substructures that contribute in a disproportionate manner. In this work, we provide a theoretical analysis of how LoRA weight updates are influenced by a layer’s activation magnitude. We propose Act-LoRA, a simple activation-guided layer selection strategy for selective Low-Rank Adaptation. We evaluate this strategy for both encoder-only and decoder-only architectures using the GLUE benchmark. Our method achieved a 20% GPUh saving with a 1% drop in GLUE score using DeBERTaV3-Base on a single-instance GPU with 50% less LoRA parameters. It also achieved 2% GPUh savings with a less than 0.15% drop in GLUE score with the Llama-3.1-8B model in Distributed Data Parallel mode with 25% fewer LoRA parameters. Our experiments and analysis show that the compute and memory requirements of LoRA adapters increase linearly with the number of selected layers. We further compare activation-guided selection against gradient-guided importance metrics and show that activation norms yield more stable and reproducible layer rankings across seeds and datasets. Overall, our results demonstrate that activation-guided layer selection is a practical and effective way to improve the efficiency of LoRA fine-tuning, making it immediately compatible with some existing PEFT techniques and distributed training frameworks. Full article

Background: This paper investigates the use of the Segment Anything Model (SAM) for chest X-ray (CXR) image segmentation, with a focus on improving its performance using low-rank adaptation (LoRA). Methods: We evaluate three versions of SAM: two zero-shot methods (using coordinate and bounding box prompts) and a fine-tuned SAM using LoRA. To support these approaches, we also trained two standard convolutional neural networks (CNNs), U-Net and DeepLabv3+, to generate draft lung segmentations that serve as input prompts for the SAM methods. Our fine-tuning approach uses LoRA to add lightweight trainable adapters within the Transformer blocks of the SAM, allowing only a small subset of parameters to be updated. The rest of the SAM remains frozen, helping preserve its pre-trained knowledge while reducing memory and computational needs. We tested all models on a dataset of CXR images labeled for COVID-19, viral pneumonia, and normal cases. Results: Results show that fine-tuned SAM with LoRA outperforms zero-shot SAM methods and CNN baselines in terms of segmentation accuracy and efficiency. Conclusions: This demonstrates the potential of combining LoRA with SAM for practical and effective medical image segmentation. Full article

While objective street metrics derived from imagery or GIS have become standard in urban analytics, they remain insufficient to capture subjective perceptions essential to inclusive urban design. This study introduces a novel Multimodal Street Evaluation Framework (MSEF) that fuses a vision transformer (VisualGLM-6B) with a large language model (GPT-4), enabling interpretable dual-output assessment of streetscapes. Leveraging over 15,000 annotated street-view images from Harbin, China, we fine-tune the framework using Low-Rank Adaptation(LoRA) and P-Tuning v2 for parameter-efficient adaptation. The model achieves an F1 score of 0.863 on objective features and 89.3% agreement with aggregated resident perceptions, validated across stratified socioeconomic geographies. Beyond classification accuracy, MSEF captures context-dependent contradictions: for instance, informal commerce boosts perceived vibrancy while simultaneously reducing pedestrian comfort. It also identifies nonlinear and semantically contingent patterns—such as the divergent perceptual effects of architectural transparency across residential and commercial zones—revealing the limits of universal spatial heuristics. By generating natural-language rationales grounded in attention mechanisms, the framework bridges sensory data with socio-affective inference, enabling transparent diagnostics aligned with Sustainable Development Goal 11(SDG 11). This work offers both methodological innovation in urban perception modeling and practical utility for planning systems seeking to reconcile infrastructural precision with lived experience. Full article

Structural damage detection using the finite element method is inherently formulated as an inverse problem, often suffering from ill-posedness and high sensitivity to measurement noise. This study introduces a novel damage detection methodology by applying low-rank adaptation (LoRA), originally developed for fine-tuning large language models, to inverse problems in structural mechanics for the first time. The proposed approach exploits the physically inherent low-rank nature of structural damage: damage is typically localized, and the contribution of each finite element to the stiffness matrix is limited by its degrees of freedom. Accordingly, the stiffness change matrix is factorized into two low-rank matrices, reducing the number of parameters and providing implicit regularization against full-rank measurement noise. Physical consistency is ensured through sparsity and symmetry constraints. Numerical experiments on cantilever beam and L-shaped plate structures across five damage scenarios demonstrated that the proposed method achieved superior noise robustness compared with baseline methods. At a signal-to-noise ratio of 20 dB, representative of practical field conditions, LoRA achieved stiffness errors below 2%, whereas the baseline methods failed to provide reliable results. The proposed framework achieved a 100% success rate in damage zone localization (Precision@n ≥ 80%) with over 60% parameter reduction, presenting a promising solution for practical structural health monitoring. Full article

High-quality text corpora are essential for knowledge graph construction and domain-specific large model pre-training in technology-intensive industries, with the steel metallurgy sector serving as a representative case. However, many industrial documents remain in scanned or PDF formats, where general-purpose Optical Character Recognition (OCR) systems exhibit systematic errors when recognizing Chinese metallurgical documents. In particular, visually similar Chinese characters that differ by only minor strokes are frequently confused, leading to severe degradation of text reliability and cascading errors in downstream knowledge extraction. This paper proposes FeOCR, a general-purpose domain-adaptive framework for machine-printed Chinese characters, which is specifically evaluated within the context of the steel metallurgy industry. The framework integrates visual character disambiguation with context-aware semantic correction. We first construct a metallurgy-specific OCR dataset emphasizing high-frequency confusable Chinese word pairs and enhance data diversity through font perturbation and noise synthesis. Parameter-efficient fine-tuning (LoRA) is then applied to adapt a general OCR model to domain-specific visual patterns. Furthermore, a Large Language Model-based correction module performs semantic refinement of residual errors under domain lexical constraints. Experiments demonstrate significant reductions in character and word error rates, especially for confusable technical terms, providing a reliable foundation for industrial Chinese document digitization. Full article

Aspect-based sentiment analysis (ABSA) plays a vital role in deriving fine-grained sentiment from textual content. As large language models (LLMs) are increasingly adopted for automated data annotation in natural language processing (NLP), concerns have emerged regarding the accuracy of their outputs. Despite their capacity to generate large volumes of labeled data, LLMs often suffer from overconfidence in predictions, high uncertainty in complex contexts, and difficulty capturing nuanced meanings, which compromise the quality of annotations and, in turn, the performance of downstream models. This underscores the need to enhance LLM adaptability while maintaining annotation accuracy. To address these limitations, we integrated low-rank adaptation (LoRA) with parameter-efficient fine-tuning (PEFT) for adapting Qwen2.5 to ABSA. LoRA reduces the number of trainable parameters by decomposing weight updates into low-rank matrices, while PEFT introduces modular adapter layers with scaled gradient updates and dynamic rank allocation. Using the standard SemEval 2014 Laptop dataset, Qwen2.5-3B fine-tuned with LoRA and PEFT achieves 64.50% accuracy, outperforming its baseline of 24.05%. Likewise, Qwen2.5-7B attains 77.50%, compared with a baseline of 34.63%. These results highlight the potential of parameter-efficient methods to improve the accuracy of LLMs in ABSA annotation tasks, especially under resource constraints. Such results lay the groundwork for scalable, reproducible LLM deployment and open avenues for future research in cross-domain adapter transferability and dynamic rank optimization. Full article

Outbreaks of stored grain pests can pose significant threats to food security. In-depth analyses of sudden outbreaks are key to achieving effective prevention and control. To address the issue of models’ insufficient reasoning capability arising from complex causal relationships in stored grain pest events, this study proposes an Event Patterns Enhancing Causal Reasoning (EPECR) method incorporating category theory. Specifically, we focus on common pests such as Sitophilus zeamais (maize weevil) and Sitotroga cerealella (Angoumois grain moth). We formally map the domain ontology—including entities like environmental factors (e.g., temperature, humidity) and control measures (e.g., fumigation)—to categories, and represent their inter-relationships (e.g., inhibition, promotion) as functors. To handle complex scenarios, we model multi-cause events (e.g., high temperature and humidity jointly accelerating pest reproduction) using functor products, and represent multi-hop events (e.g., environmental changes leading to pest outbreak and subsequent grain loss) through functor compositions. This formal expression enables Large Language Models (LLMs) to extract reliable event patterns. Based on these patterns, this study constructed 1440 structured datasets and adopted the Low-Rank Adaptation (LoRA) strategy to fine-tune the LLMs. Experiments on the domain-specific Stored Grain Pest Events Dataset (SGPE) demonstrate that EPECR achieves a reasoning accuracy of 85.9% on in-distribution data and 79.9% on out-of-distribution data, effectively identifying correct causal chains for pest logic. This method significantly outperforms the state-of-the-art domain method-Naive Augmentations (NA)-by 4.9%, providing precise decision support for the early warning and control of specific pest incidents. Full article