This paper addresses key challenges in knowledge injection for large language models, including static representation, difficulty in updating, and limited domain adaptability. It proposes a dynamic fine-tuning framework designed for knowledge injection. The framework is based on parameter-efficient tuning strategies and introduces learnable adapter modules and gating mechanisms. These components enable selective integration and dynamic control of both structured and unstructured external knowledge. In the model design, a query encoder extracts semantic vectors from the input text. These vectors are matched with an external knowledge base to construct a dynamic knowledge subset. This subset guides task generation. Adapter modules and gating units are then applied across model layers to adjust knowledge enhancement. This ensures that external knowledge contributes to reasoning while preserving the model's original language capabilities. A unified joint loss function is also introduced. It coordinates the optimization between language modeling and knowledge alignment objectives. To evaluate the proposed method, the WikiHop dataset for multi-hop question answering is used. The model's behavior is analyzed under various experimental settings, including different parameter update ratios, knowledge densities, and cross-domain transfer scenarios. The results show that the method achieves strong performance on key metrics such as knowledge recall, F1 score, and inference efficiency, even with low parameter update ratios. This demonstrates the practicality and stability of the approach in dynamic knowledge integration tasks. The proposed work offers a flexible and efficient technical path for knowledge injection and domain adaptation in large language models.

Large Language Model (LLM)-based multi-agent systems have emerged as a promising paradigm for tackling complex tasks that exceed individual agent capabilities. However, existing approaches often suffer from coordination inefficiencies, a lack of trust mechanisms, and suboptimal role assignment strategies. This paper presents a novel trust-aware coordination framework that enhances multi-agent collaboration through dynamic role assignment and context sharing. Our framework introduces a multi-dimensional trust evaluation mechanism that continuously assesses agent reliability based on performance history, interaction quality, and behavioral consistency. The coordinator leverages these trust scores to dynamically assign roles and orchestrate agent interactions while maintaining a shared context repository for transparent information exchange. We evaluate our framework across eight diverse task scenarios with varying complexity levels, demonstrating significant improvements over baseline approaches. Experimental results show that our trust-aware framework achieves a 87.4% task success rate, reducing execution time by 36.3% compared to non-trust-based methods, while maintaining 43.2% lower communication overhead. The framework's ability to adapt agent roles based on evolving trust scores enables more efficient resource utilization and robust fault tolerance in dynamic multi-agent environments.

Detecting anatomical tooth landmarks on intraoral 3D scans (IOS) is a key step in digital orthodontics, but supervised methods are limited by the few scans carrying landmark annotations. On the public Teeth3DS+ benchmark every scan has per-vertex tooth segmentation and FDI labels, whereas only a small subset also carries dental landmarks, so landmark detection becomes a semi-supervised problem in which abundant segmentation labels compensate for the missing landmark labels. We pretrain a shared point-cloud encoder on segmentation using all non-test scans, then train a landmark head with an equivariance consistency loss and confidence-filtered pseudo-labels on the unlabelled scans, plus a geometric prior that constrains each predicted landmark to the segmented region of its tooth. All model selection uses a held-out validation split; the test set serves only for final reporting. On the 3DTeethLand test split the method raises the mean Average Precision from 0.702 to 0.753 and lowers the mean radial error from 0.63 to 0.54 mm over a supervised baseline with the same labelled budget, ranking third of seven systems. Existing segmentation labels are thus an effective, low-cost signal for label-efficient 3D landmark detection.

Multi-agent systems (MAS) have emerged as a critical paradigm for distributed problem-solving in complex environments. However, their deployment in mission-critical applications faces significant challenges regarding trust, security, and adversarial robustness. This paper presents TrustOrch, a novel dynamic trust-aware orchestration framework designed to enhance the resilience of multi-agent collaboration against adversarial attacks. TrustOrch introduces five key innovations: (1) a dynamic trust assessment mechanism that evaluates agent reliability in real-time using multi-dimensional metrics, (2) an adversary-aware orchestration strategy combining reinforcement learning and game theory to detect and mitigate prompt injection attacks, (3) an adaptive collaboration topology that dynamically adjusts agent communication structures based on task complexity and trust levels, (4) explainable decision tracing for complete audit chains, and (5) a layered security architecture leverag- ing blockchain technology for decentralized trust verification. Our experimental evaluation demonstrates that TrustOrch re- duces collision rates by 62%, achieves 91.7% robustness under adversarial attacks, and reduces communication overhead by 39.8% compared to baseline approaches. The framework achieves robust performance under various adversarial scenarios while maintaining transparency and regulatory compliance, making it particularly suitable for deployment in high-risk domains such as finance, healthcare, and autonomous systems.

Breast cancer is a relatively common cancer among gynecological cancers. Its diagnosis often relies on the pathology of cells in the lesion. The pathological diagnosis of breast cancer not only requires professionals and time, but also sometimes involves subjective judgment. To address the challenges of dependence on pathologists expertise and the time-consuming nature of achieving accurate breast pathological image classification, this paper introduces an approach utilizing convolutional neural networks (CNNs) for the rapid categorization of pathological images, aiming to enhance the efficiency of breast pathological image detection. And the approach enables the rapid and automatic classification of pathological images into benign and malignant groups. The methodology involves utilizing a convolutional neural network (CNN) model leveraging the Inceptionv3 architecture and transfer learning algorithm for extracting features from pathological images. Utilizing a neural network with fully connected layers and employing the SoftMax function for image classification. Additionally, the concept of image partitioning is introduced to handle high-resolution images. To achieve the ultimate classification outcome, the classification probabilities of each image block are aggregated using three algorithms: summation, product, and maximum. Experimental validation was conducted on the BreaKHis public dataset, resulting in accuracy rates surpassing 0.92 across all four magnification coefficients (40X, 100X, 200X, and 400X). It demonstrates that the proposed method effectively enhances the accuracy in classifying pathological images of breast cancer.

In warehouse bin picking, a manipulator must decide which object to remove first, because grasping a pinned object often fails or drags neighbouring items out of the bin. Prevailing approaches rely on amodal instance segmentation: the full extent of each partially occluded object is first predicted, and occlusion is inferred from the area ratio between the visible and the completed mask. Completing occluded regions is ill-posed and multi-modal, so shape errors propagate into the pick order while incurring substantial computational cost. We propose a more direct formulation: pick-order reasoning does not require hidden shape, only one visibility-ratio scalar per instance. We attach a lightweight regression head to a frozen segmentation backbone and predict the ratio of visible to complete area directly from instance features. Supervision is obtained at no additional cost from the amodal annotations of existing datasets. We evaluate visibility error, rank correlation and top-layer retrieval on three public benchmarks, and first-grasp success on a physical manipulator. The proposed method yields pick orders on par with representative amodal pipelines while running roughly five to seven times faster than their released implementations.