The rapid evolution of Large Language Models has revolutionized natural language processing, yet these models frequently exhibit limitations when deployed in specialized high-stakes domains such as medicine, law, and engineering. A primary deficiency is the propensity for hallucination and the inability to access up-to-date, structured factual knowledge that was not present or emphasized during the pre-training phase. This paper proposes a novel architecture that integrates Domain-Specific Knowledge Graphs with pre-trained language models utilizing Graph Attention Networks. By employing a dual-stream mechanism that processes textual input alongside structured graph data, we facilitate a deep injection of semantic relationships into the latent space of the language model. The Graph Attention Network component dynamically weighs the importance of neighboring entities within the knowledge graph, allowing the model to attend to the most relevant factual context corresponding to the input query. We evaluate this approach on two distinct domain-specific datasets involving biomedical and legal texts. Our experimental results demonstrate that this injection mechanism significantly outperforms standard fine-tuning approaches in terms of factual accuracy and reasoning capabilities. The proposed method offers a scalable pathway toward creating more reliable and logically sound domain-specific artificial intelligence systems.

Abstract Large-language models (LLMs) have transformed natural language processing and opened new possibilities for the computational social sciences and digital humanities. Yet translating historical sources remains difficult because early modern varieties are scarcely represented in contemporary training corpora and because standard tokenizers fragment their non-standard orthography. This article tackles these gaps by adapting open LLMs to early modern Dutch-to-English translation and advances two concrete contributions: (i) a memory-efficient fine-tuning workflow that runs on a single consumer GPU, comparing order-reward policy optimization with the Unsloth supervised fine-tuning approach and (ii) a verifiable evaluation protocol that combines embedding-based metrics with systematic expert review. Experiments on testimonial texts (1680–1792) show that fine-tuning choice decisively shapes quality: the Unsloth-tuned Mistral model attains the highest BERTScore and METEOR values and most faithfully preserves historical nuance. The framework supports a collaborative workflow where machine-generated drafts accelerate expert translation, making archival texts more accessible while maintaining scholarly oversight through domain-expert validation.

Within the realm of software engineering, specialized tasks on code, such as program repair, present unique challenges, necessitating fine-tuning Large language models (LLMs) to unlock state-of-the-art performance. Fine-tuning approaches proposed in the literature for LLMs on program repair tasks generally overlook the need to reason about the logic behind code changes, beyond syntactic patterns in the data. High-performing fine-tuning experiments also usually come at very high computational costs. With MORepair , we propose a novel perspective on the learning focus of LLM fine-tuning for program repair: we not only adapt the LLM parameters to the syntactic nuances of the task of code transformation (objective ➊), but we also specifically fine-tune the LLM with respect to the logical reason behind the code change in the training data (objective ➋). Such a multi-objective fine-tuning will instruct LLMs to generate high-quality patches. We apply MORepair to fine-tune four open-source LLMs with different sizes and architectures. Experimental results on function-level and repository-level repair benchmarks show that the implemented fine-tuning effectively boosts LLM repair performance by 11.4% to 56.0%. We further show that our fine-tuning strategy yields superior performance compared to the state-of-the-art approaches, including standard fine-tuning, Fine-tune-CoT, and RepairLLaMA.

Dementia is a progressive neurodegenerative disease that is a major challenge to healthcare systems around the world and hence there is the need to have accurate and automated means of diagnosis. SMRI has become a useful modality for detecting neuroanatomical changes during the development of dementia, and yet, manual interpretation is tedious and prone to error. This paper will discuss a deep learning-based method of multiclass dementia classification based on a transfer learning model on the VGG-16 convolutional neural network. A big publicly available Kaggle data set comprising of about 44,000 T1-weighted brain MRI images was used, which comprised of four clinically relevant classes such as Non-Demented, Very Mild Demented, Mild Demented as well as Moderate Demented. All pictures were downsampled to the constant resolution size at 224 x 224 pixels and categorized as grayscale inputs to retain structural information and making them computationally efficient. A fine-tuning approach that was under a controlled strategy was used by unfreezing convolutional layers consecutively, allowing the detailed examination of convolutional layers parameter adaptation and generalization behaviours. The evaluation of the model performance was conducted based on accuracy metrics, learning curves, and analysis of the confusion matrix to present both quantitative and class-wise information. The final training accuracy of the proposed model was 89 percent and a validation accuracy of 76 percent which showed that the model converged well and the generalization was also good. The confusion matrix showed that Non-Demented cases were highly specific with potential difficulties likely to arise in making a distinction between the early stages of dementia since there were minor neuroanatomical overlaps.

Dental plaques are biofilms formed by microorganisms and extracellular matrix in the oral cavity. The most common oral diseases, including dental caries and periodontal diseases, are initiated by dental plaque accumulating on tooth surfaces. We develop an AI-based algorithm (AI-Plaque) to automatically segment dental plaque areas in intraoral images, enabling automatic plaque detection and quantitative severity assessment. To mitigate the limited segmentation capabilities of existing models and dataset limitations, we introduce a data preprocessing procedure, a specialized fine-tuning approach, and a novel self-training pipeline that leverages synthesized plaque images. Our approach demonstrates significant improvements in plaque segmentation and high effectiveness in plaque assessment, over several baseline methods, as evaluated by both image segmentation metrics and a custom-designed quantitative Visual Plaque Index. The AI-Plaque algorithm could empower individuals to receive timely, personalized feedback on their oral hygiene practices for dental plaque control, such as brushing and flossing, and ultimately reduce the risk of developing oral diseases.

Adenosine-to-inosine (A-to-I) RNA editing by ADAR enzymes shapes transcript fate and underpins emerging RNA editing therapeutics, yet predicting which adenosines are edited remains difficult. We introduce ADAR-GPT, a model-agnostic fine-tuning framework that adapts a GPT-class language model to classify editing at candidate sites using sequence context in standardized 201 nt windows with the target adenosine explicitly marked. We train and evaluate on GTEx liver data ([Formula: see text] samples) at a clinically relevant 15% editing threshold, using a two-stage continual fine-tuning approach where lower thresholds serve as curriculum data to progressively sharpen decision boundaries. Using sequence data, ADAR-GPT demonstrates competitive or superior performance when benchmarked against established computational approaches, including convolutional and foundation model architectures, achieving a better balance of recall, precision, and specificity alongside stronger operating-curve metrics. The approach is reproducible and portable across GPT backbones without architectural changes. Beyond accurate site classification, ADAR-GPT provides practical adenosine scoring to prioritize experimental targets and inform guide RNA design, with a framework adaptable to new datasets and model architectures.

Recent advances in spatial omics technologies have enabled transcriptome profiling at subcellular resolution. By performing cell segmentation on nuclear or membrane staining images, researchers can acquire single cell level spatial gene expression data, which in turn enables subsequent biological interpretation. Although deep learning-based segmentation models achieve high overall accuracy, their performance remains suboptimal for whole-tissue analysis, particularly in ensuring consistent segmentation accuracy across diverse cell populations. Existing fine-tuning approaches often require extensive retraining or are tailored to specific model architectures, limiting their adaptability and scalability in practical settings. To address these challenges, we present CSRefiner, a lightweight and efficient fine-tuning framework for precise whole-tissue single-cell spatial expression analysis. Our approach incorporates support for fine-tuning widely used segmentation models in the field of spatial omics, while achieving high accuracy with very limited annotated data. This study demonstrates CSRefiner’s superior performance across various staining types and its compatibility with multiple mainstream models. Combining operational simplicity with robust accuracy, our framework offers a practical solution for real-world spatial transcriptomics applications.

Direct ethanol fuel cells (DEFCs) have attracted much attention due to their high energy density and readily available fuels, but their development is limited by the poor performance of anode catalysts. Although Pd-based catalysts exhibit excellent activity in the ethanol oxidation reaction (EOR), their active site density is low, and they are easily poisoned by strong adsorption of CO* intermediates. Herein, we designed unique nanoflower-like AuPdCu nanoparticles (NPs) through a liquid-phase reflux process. Compared to conventional nanoflowers, their petals consist of numerous converging nanowires. Furthermore, the electronic structure of Pd-based catalysts can be finely tuned by alloying with Au and Cu to regulate the adsorption and desorption of CO* intermediates. As a result, the designed AuPdCu NPs demonstrated exceptional catalytic activity for the EOR, achieving a mass activity (MA) of 6.89 A mgPd-1, surpassing commercial Pd/C and some recently reported catalysts. After a 5000 s chronoamperometric stability test, they maintained a current density of 34.81 mA cm-2. Density functional theory (DFT) calculations confirm that the adsorption of the CH3CO* intermediate on AuPdCu NPs is enhanced, thereby promoting the EOR process along the C1 pathway. This ternary metal fine-tuning alloying approach presents a viable route for fabricating highly active and durable EOR materials.

With the increasing frequency of port trade activities, oil spill incidents pose severe threats to marine ecosystems. The complex port environments, characterized by diverse land cover types and intricate land-water interfaces, present significant challenges for oil spill detection (OSD). To address this, we propose a parameter-efficient fine-tuning (PEFT) approach based on the segment anything model (SAM) for port OSD tasks utilizing unmanned aerial vehicle (UAV) remote sensing imagery. The proposed DADS-SAM freezes most parameters of SAM and performs PEFT only on adapters. Dynamic adaptive low-rank adaptation (DALoRA) enhances the model’s adaptability to port OSD tasks by incorporating dynamic adaptive low-rank matrices into the self-attention modules of the SAM encoder, while dual scale convolutional adapter (DSCA) improves the model’s capability to capture local features and spatial dependencies in oil spill regions. Since the original SAM model lacks class-specific segmentation capabilities, we further design a multi-class segmentation head (MCSH) for oil spill segmentation. This head obtains the output masks from the SAM decoder, adjusts their dimensions, and transforms them through two linear layers to generate class-specific segmentation results. To the best of our knowledge, this study is the first to apply a PEFT approach to port OSD tasks. The proposed method was evaluated on both publicly available port oil-spill datasets and marine oil-spill datasets acquired via synthetic-aperture radar (SAR). Quantitative comparisons demonstrate that DADS-SAM outperforms conventional approaches and achieves state-of-the-art performance across all evaluated metrics. Models and data are publicly available at https://github.com/PetrichorAsh/DADS-SAM.

The rapid growth of model scale has necessitated substantial computational resources for fine-tuning. Existing approach such as low-rank adaptation (LoRA) has sought to address the problem of handling the large updated parameters in full fine-tuning (FT). However, LoRA utilize random initialization and optimization of low-rank matrices to approximate updated weights, which can result in suboptimal convergence and an accuracy gap compared to full fine-tuning (FT). To address these issues, we propose low-rank LDU (LoLDU), a parameter-efficient fine-tuning (PEFT) approach that significantly reduces trainable parameters by 2600 times compared to regular PEFT methods while maintaining comparable performance. LoLDU leverages lower-diag-upper (LDU) decomposition to initialize low-rank matrices for faster convergence and nonsingularity. We focus on optimizing the diagonal matrix for scaling transformations. To the best of our knowledge, LoLDU has the fewest parameters among all PEFT approaches. We conducted extensive experiments across 4 instruction-following datasets, six natural language understanding (NLU) datasets, eight image classification datasets, and image generation datasets with multiple model types [LLaMA2, RoBERTa, ViT, and stable diffusion (SD)], providing a comprehensive and detailed analysis. Our open-source code can be accessed at https://anonymous.4open.science/r/LoLDU-B5A6.

Photovoltaic power forecasting based on large language models: A prompt-learning and two-stage fine-tuning approach

The malicious misuse and widespread dissemination of AI-generated images pose a significant threat to the authenticity of online information. Current detection methods often struggle to generalize to unseen generative models, and the rapid evolution of generative techniques continuously exacerbates this challenge. Without adaptability, detection models risk becoming ineffective in real-world applications. To address this critical issue, we propose a novel three-stage domain continual learning framework designed for continuous adaptation to evolving generative models. In the first stage, we employ a strategic parameter-efficient fine-tuning approach to develop a transferable offline detection model with strong generalization capabilities. Building upon this foundation, the second stage integrates unseen data streams into a continual learning process. To efficiently learn from limited samples of novel generated models and mitigate overfitting, we design a data augmentation chain with progressively increasing complexity. Furthermore, we leverage the Kronecker-Factored Approximate Curvature (K-FAC) method to approximate the Hessian and alleviate catastrophic forgetting. Finally, the third stage utilizes a linear interpolation strategy based on Linear Mode Connectivity, effectively capturing commonalities across diverse generative models and further enhancing overall performance. We establish a comprehensive benchmark of 27 generative models, including GANs, deepfakes, and diffusion models, chronologically structured up to August 2024 to simulate real-world scenarios. Extensive experiments demonstrate that our initial offline detectors surpass the leading baseline by +5.51% in terms of mean average precision. Our continual learning strategy achieves an average accuracy of 92.20%, outperforming state-of-the-art methods.

Accurate segmentation of organs-at-risk is critical for adaptive Magnetic Resonance guided radiotherapy (MRgRT) in upper gastrointestinal (GI) treatments. Large anatomical variations pose challenges for automated segmentation methods, requiring considerable manual editing. Our goal was to develop and evaluate the effectiveness of patient-specific (PS) fine-tuning of a pretrained transformer model for segmenting upper GI structures. A PS fine-tuning approach was implemented by adapting a pretrained transformer model called Self-distilled Masked Image Transformer (SMIT). SMIT was trained using T2-weighted MR images from 30 patients with pancreatic cancers who underwent MRgRT to create a general model (GM). This GM was subsequently adapted to each new individual patient using data from their first treatment fraction, producing a PS model. Its performance was evaluated across the subsequent treatment fractions in 10 patients using Dice similarity coefficient (DSC), Hausdorff distance (HD95), and added path length (APL) by comparing against expert delineations. PS model outperformed the GM across all structures, with the largest improvements for duodenum-stomach (DSC: 0.84 to 0.90, HD95: 15.9 to 4.9 mm). This approach reduced potential contour editing efforts, decreasing median APL by 4%, 7% and 26% for large bowel, small bowel, and duodenum-stomach. PS fine-tuning was most effective for scans with low soft-tissue contrast and challenging anatomy. This method provides a practical, computationally feasible approach to enhance online adaptive workflow by improving segmentation accuracy and potentially reducing the need for manual editing that overall improves on-table treatment times.

In this paper, we introduce a simple but effective training-free pipeline for handling the task of text-to-video object segmentation. Our approach leverages open-source Multimodal Large Language Models (MLLMs) for segmenting objects in videos based on language descriptions. We design three multimodal reasoning agents that decompose the task into semantic, temporal, and spatial reasoning stages: a Video Summarization Agent to provide concise semantic context, a Keyframe Selection Agent employing a Binary-Logit Frame Scoring mechanism to identify informative frames, and an Object Grounding Agent predicting bounding boxes for the described objects. Finally, by providing high-quality prompts to a semantic-free segmentation tool, our method effectively handles spatiotemporal variations and reduces segmentation errors. Extensive experiments show that our training-free method significantly reduces resource requirements while achieving comparable or even better performance than supervised fine-tuning approaches.

Machine learning models, especially neural networks, are vulnerable to adversarial attacks, where inputs are purposefully altered to induce incorrect predictions. These adversarial inputs closely resemble benign (unaltered) inputs, making them difficult to detect, and pose significant security risks in critical applications, such as autonomous vehicles, medical diagnostics, and financial transactions. Several methods exist to improve the model’s performance against these adversarial attacks, which typically modify the network architecture or training procedure. Often times, these adversarial training techniques only provide robustness against specific attack types and/or require substantial computational resources, making them impractical for real-world applications with limited resources. In this work, we propose a computationally-efficient adversarial fine-tuning approach to enhance the robustness of Convolutional Neural Networks (CNNs) against adversarial attacks and attain the same level of performance as the conventional adversarial training. More specifically, we propose to identify specific parts of the neural network model that are more vulnerable to adversarial attacks. Our analysis reveals that only a small portion of these vulnerable components accounts for a majority of the model’s errors caused by adversarial attacks. As such, we propose to selectively fine-tune these vulnerable components using different adversarial training methods to develop an effective and resource-efficient approach to improve model robustness. We empirically validate our proposed approach with varying dataset and algorithm parameters. We demonstrate that our approach can achieve similar performance as the more resource-intensive conventional adversarial training method. Received: 18 April 2025 | Revised: 4 September 2025 | Accepted: 12 November 2025 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in github.io at https://doi.org/10.1109/5.726791, reference number [3]. Author Contribution Statement Subah Karnine: Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization. Sadia Afrose: Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization. Hafiz Imtiaz: Conceptualization, Methodology, Formal analysis, Resources, Writing – original draft, Writing – review & editing, Supervision, Project administration.

Recent breakthroughs in large-scale speech models such as Whisper, Wav2Vec 2.0, and HuBERT have greatly enhanced speech processing tasks. Full fine-tuning comes at a prohibitive cost, though, which restricts their application to low- resource or real-time settings. The parameter-efficient fine-tuning (PEFT) approaches—e.g., LoRA, QLoRA, adapters, and prompt tuning—allow for compact adaptation by fine-tuning only a small subset of parameters. We review 33 studies (2021–2025) using PEFT for applications such as ASR, speaker verification, and emotion recognition. We organize methods by task, compare efficiency and accuracy, and determine prominent trends. Results indicate PEFT produces competitive results with reduced cost, enabling scalable deployment in resource-poor environments.

BackgroundAneurysmal subarachnoid hemorrhage (SAH) is a life-threatening neurological emergency with mortality rates exceeding 30%. While deep learning techniques show promise for automated SAH segmentation, their clinical application is limited by the scarcity of labeled data and challenges in cross-institutional generalization. Transfer learning from related hematoma types represents a potentially valuable but underexplored approach. Although Unet architectures remain the gold standard for medical image segmentation due to their effectiveness on limited datasets, Low-Rank Adaptation (LoRA) methods for parameter-efficient transfer learning have been rarely applied to convolutional neural networks in medical imaging contexts. The importance of SAH diagnosis and the time-intensive nature of manual annotation would benefit from automated solutions that can leverage existing multi-institutional datasets from more common conditions.MethodsWe implemented a Unet architecture pre-trained on computed tomography scans from 124 traumatic brain injury patients across multiple institutions, then fine-tuned on 30 aneurysmal SAH patients from the University of Michigan Health System using 3-fold cross-validation. We developed a novel CP-LoRA method based on tensor canonical polyadic (CP) decomposition and introduced DoRA variants (DoRA-C, convDoRA, CP-DoRA) that decompose weight matrices into magnitude and directional components. We compared these approaches against existing LoRA methods (LoRA-C, convLoRA) and standard fine-tuning strategies across different modules on a multi-view Unet model. Performance was evaluated using Dice scores stratified by hemorrhage volume, with additional assessment of predicted versus annotated blood volumes.ResultsTransfer learning from traumatic brain injury to aneurysmal SAH demonstrated feasibility with all fine-tuning approaches achieving superior performance compared to no fine-tuning (mean Dice 0.410 ± 0.26). The best-performing traditional approach was decoding module fine-tuning (Dice 0.527 ± 0.20). LoRA-based methods consistently outperformed standard Unet fine-tuning, with DoRA-C at rank 64 achieving the highest overall performance (Dice 0.572 ± 0.17). Performance varied by hemorrhage volume, with all methods showing improved accuracy for larger volumes (Dice 0.682–0.694 for volumes > 100 mL vs. Dice 0.107–0.361 for volumes < 25 mL). CP-LoRA achieved comparable performance to existing methods while using significantly fewer parameters. Over-parameterization with higher ranks (64–96) consistently yielded better performance than strictly low-rank adaptations.ConclusionsThis study demonstrates that transfer learning between hematoma types is feasible and that LoRA-based methods significantly outperform conventional Unet fine-tuning for aneurysmal SAH segmentation. The novel CP-LoRA method offers parameter efficiency advantages, while DoRA variants provide superior segmentation accuracy, particularly for small-volume hemorrhages. The finding that over-parameterization improves performance challenges traditional low-rank assumptions and suggests clinical applications may benefit from higher-rank adaptations. These results support the potential for automated SAH segmentation systems that leverage large multi-institutional traumatic brain injury datasets, potentially improving diagnostic speed and consistency when specialist expertise is unavailable.

Although Convolutional Neural Networks (CNNs) have delivered state-of-the-art accuracy in plant disease classification, their deployment is still hindered by data scarcity, computational cost, and architectural heterogeneity. Transfer learning from large-scale pre-trained datasets alleviates these issues, yet generic feature extraction suffers from domain shift, while indiscriminate fine-tuning risks over-fitting and elevated training budgets. To address the identified limitations, PlantClassiNet is implemented as a unified framework. This framework facilitates systematic comparative analysis of six CNN architectures, AlexNet, ResNet50, InceptionV3, MobileNetV3Small, DenseNet121 and EfficientNetB0, across three publicly available datasets: PlantVillage, PlantLeaves and Eggplant. Two alternative fine-tuning approaches are proposed: ‌Adaptive Fine-tuning (AdapFitu)‌, which adaptively determines the depth of unfrozen layers, learning rates, and reinitializes selected layers, and a ‌fixed-parameter baseline‌, which trains only the newly added classifier while keeping the convolutional backbone frozen and unfreezes a fixed number of network layers for retraining. Extensive experiments demonstrate that large models AlexNet, ResNet50, and Inceptionv3 achieve test accuracy exceeding 98.74% on the sizable PlantVillage dataset, whereas lightweight counterparts MobileNetV3Small, DenseNet121, and EfficientNetB0 achieve high accuracy of 99.79% ± 0.21% on the smaller Eggplant collection after fine-tuning.

This study aims to establish a method-integrative framework for emotion-oriented architectural image generation. The framework combines Stable Diffusion with targeted LoRA (Low-Rank Adaptation), a lightweight and parameter-efficient fine-tuning approach, together with ControlNet-based structural constraints, to examine how controllable design-element manipulations influence emotional responses. The methodology follows a closed-loop “generation–evaluation” workflow, with each LoRA module independently targeting a single design element. Guided by the relaxation–arousal emotional dimension, the framework is evaluated using subjective ratings and electroencephalogram (EEG) measures. Twenty-seven participants viewed six architectural space categories, each comprising four conditions (baseline, color, material, and form modification). EEG α/β power ratio (RAB) served as the primary neurophysiological marker of arousal. Statistical analysis indicated that LoRA-based modifications of design elements produced distinct emotional responses: color and material changes induced lower arousal, whereas changes in form elicited a bidirectional pattern involving relaxation and arousal. The right parietal P4 electrode site showed the most sensitive emotional response to design element changes, with consistent statistical significance. P4 is a human scalp EEG location associated with cortical activity related to visuospatial processing. Descriptive results suggested opposite directional effects with similar intensity trends; however, linear mixed-effects model (LMM) inference did not support significant group-level linear coupling, indicating individual variation. This study demonstrates the feasibility of emotion-guided architectural image generation, showing that controlled manipulation of color, material, and form can elicit measurable emotional responses in human brain activity. The findings provide a methodological basis for future multimodal, adaptive generative systems and offer a quantitative pathway for investigating the relationship between emotional states and architectural design elements.

ABSTRACT 1. Early identification of various poultry illnesses may be possible with deep learning techniques. Salmonellosis, newcastle disease and coccidiosis illnesses can present poor faecal scores as symptoms. A deep convolutional neural network (CNN) model known as an ‘ensemble-based CNN’ was constructed to identify poultry diseases by classifying healthy and unhealthy faecal images and tested on farm. 2. Four distinct deep learning models; MobileNetV2, EfficientNetB0, Xception and ResNet50V2, were compared and their corresponding inference findings recorded. To identify the best results, ensemble techniques and transfer learning, using a fine-tuning approach, were applied. 3. Evaluation scores for accuracy, precision and recall were used to assess the four models to determine which performed best. 4. The sum-ensemble approach and weighted average ensemble algorithm achieved 99% and the majority voting ensemble achieved 98.5% accuracy, respectively. 5. The Xception model outperformed the three other models in identifying type of disease. Therefore, Xception was deemed the preferred model when accuracy is the main concern whereas the ensemble approach enhanced the performance of identification.