Adapting Large Language Models (LLMs) to specialized domains like geriatric care remains a significant challenge due to the limited availability of domain-specific data and the difficulty of achieving efficient yet effective fine-tuning. Current methods often fail to effectively harness domain-specific causal insights, which are crucial for understanding and solving complex problems in low-resource domains.To address these challenges, we propose Causality-Aware, Parameter-Efficient Learning (CPEL), a novel framework that leverages domain-specific causal relationships to guide a multi-layer, parameter-efficient fine-tuning process for more effective domain adaptation. By embedding causal reasoning into the model’s adaptation pipeline, CPEL enables efficient specialization in the target domain while maintaining strong task-specific performance. Specifically, the Causal Prompt Generator of CPEL extracts and applies domain-specific causal structures, generating adaptive prompts that effectively guide the model’s learning process. Complementing this, the MPEFT module employs a dual-adapter mechanism to balance domain-level adaptation with downstream task optimization. This cohesive design ensures that CPEL achieves resource efficiency while capturing domain knowledge in a structured and interpretable manner. Based on this framework, we delved into its application in the field of geriatric care and trained a specialized large language model (Geriatric Care LLaMA) tailored for the aged-care domain, leveraging its capacity to efficiently integrate domain expertise. Experimental results from question-answering tasks demonstrate that CPEL improves ROUGE scores by 9–14% compared to mainstream LLMs and outperforms frontier models by 1–2 points in auto-scoring tasks. In summary, CPEL demonstrates robust generalization and cross-domain adaptability, highlighting its scalability and effectiveness as a transformative solution for domain adaptation in specialized, resource-constrained fields.

Accurate detection of tobacco plants in complex field environments is critical for precision agriculture, crop monitoring, and yield estimation. Traditional manual counting methods are time-consuming, labor-intensive, and susceptible to environmental and subjective factors. In this study, we propose an improved YOLO11-based framework for automated tobacco plant detection, specifically designed to address challenges such as scale variation, dense distribution, and background interference. The framework integrates four key modules: the Edge-Enhanced Feature Stem (EEFS) to strengthen low-level feature extraction, the Multi-Scale Kernel Interaction (MSKI) to capture multi-scale contextual information, the Adaptive Weighted Feature Fusion (AWFF) to optimize feature aggregation, and the Global–Local Synergistic Attention (GLSA) to enhance feature discrimination by jointly modeling local details and global context. A comprehensive UAV-based tobacco dataset was constructed, encompassing multiple lighting conditions, collection heights, and observation angles. Experimental results demonstrate that the proposed method significantly outperforms the YOLO11 baseline and achieves superior performance compared to mainstream YOLO variants. Ablation studies and heatmap visualizations confirm the effectiveness of each module. Furthermore, the model exhibits robust performance under multi-dimensional environmental perturbations, including varying illumination, scale, and camera angles. The proposed framework provides a practical and efficient solution for automated tobacco plant counting, offering potential applications in UAV-based precision agriculture and large-scale crop monitoring.

Generalized zero-shot learning (GZSL) aims to simultaneously recognize both seen classes and unseen classes by training only on seen class samples and auxiliary semantic descriptions. Recent state-of-the-art methods infer unseen classes based on semantic information or synthesize unseen classes using generative models based on semantic information, all of which rely on the correct alignment of visual–semantic features. However, they often overlook the inconsistency between original visual features and semantic attributes. Additionally, due to the existence of cross-modal dataset biases, the visual features extracted and synthesized by the model may also mismatch with some semantic features, which could hinder the model from properly aligning visual–semantic features. To address this issue, this paper proposes a GZSL framework that enhances the consistency of visual–semantic features using a self-distillation and disentanglement network (SDDN). The aim is to utilize the self-distillation and disentanglement network to obtain semantically consistent refined visual features and non-redundant semantic features to enhance the consistency of visual–semantic features. Firstly, SDDN utilizes self-distillation technology to refine the extracted and synthesized visual features of the model. Subsequently, the visual–semantic features are then disentangled and aligned using a disentanglement network to enhance the consistency of the visual–semantic features. Finally, the consistent visual–semantic features are fused to jointly train a GZSL classifier. Extensive experiments demonstrate that the proposed method achieves more competitive results on four challenging benchmark datasets (AWA2, CUB, FLO, and SUN).