AI-assisted coding tools powered by Code Large Language Models (CodeLLMs) are increasingly integrated into modern software development workflows. To address concerns around privacy, latency, and model customization, many enterprises opt to self-host these models. However, the diversity and growing number of CodeLLMs, coupled with limited accelerator memory, introduce practical challenges in model management and serving efficiency. This paper presents CACE, a novel context-aware model eviction strategy designed specifically to optimize self-hosted CodeLLM serving under resource constraints. Unlike traditional eviction strategies based solely on recency (e.g., Least Recently Used), CACE leverages multiple context-aware factors, including model load time, task-specific latency sensitivity, expected output length, and recent usage and future demand tracked through a sliding window. We evaluate CACE using realistic workloads that include both latency-sensitive code completion and throughput-intensive code reasoning tasks. Our experiments show that CACE reduces Time-to-First-Token (TTFT) by 70% and end-to-end (E2E) latency by 37%, while significantly lowering the number of model evictions by 55% compared to state-of-the-art systems. Ablation studies further demonstrate the importance of multi-factor eviction in balancing responsiveness and resource efficiency. This work contributes practical strategies for deploying scalable, low-latency AI coding assistants in real-world software engineering environments.

For a limited scale self-organized multi-agent system operating in environments with unknown task distributions, one challenge is to reduce the task response time via efficiently combining task exploration and allocation, another challenge is to improve the task completion rate via unlocking the potential of network cooperation in task allocation. However, in the existing studies, task allocation is generally regarded as an independent issue for known task distribution environments, rarely combined with task exploration, also hardly solving the conflict between the multi-hop network cooperation and mobility flexibility of agents. In view of this, this paper proposes a network connectivity-aware deep reinforcement learning method for task exploration and allocation in limited scale multi-agent systems (NCADRL4TEA). This method divides the task environment into regions and integrates task exploration with task allocation via two policies: a leaving policy to guide global task exploration among regions according to the distribution of agents and tasks, and a stay policy to guide local task allocation within each region according to the multi-hop network cooperation performance between agents. Further, in the stay policy, a network connectivity-aware task allocation optimization model is provided, which leads agents in the same region to cooperate with each other via multi-hop intermittent network connectivity and flexibly adjust their locations until the optimal multi-hop network cooperation performance is achieved. The experimental results verify that NCADRL4TEA can reduce the task response time in combination of task exploration and allocation, and improve the task completion rate in network cooperation.

Epilepsy affects millions of people worldwide and poses significant challenges for diagnosis and treatment. While traditional EEG analysis remains essential for seizure detection, it is often time-intensive and requires specialized expertise. To overcome these limitations, we proposed EpilepsyNet, an innovative deep learning architecture for automatic seizure detection. EpilepsyNet incorporates a self-calibration mechanism that enhances feature extraction by expanding the receptive field and enabling dynamic interactions across multiple EEG channels. Using neonatal EEG data, we validated EpilepsyNet’s effectiveness through five-fold cross-validation, demonstrating high classification accuracy and robustness across various performance metrics. Ablation experiments further underscored the important role of the self-calibrated convolution, as removing this component led to a notable decline in accuracy. These findings suggest that EpilepsyNet provides a scalable and reliable solution for seizure detection, with the potential to significantly improve clinical interventions by automating and optimizing the analysis of EEG data.

The web is a rich information repository that can be mined to uncover additional data about past flash flood (FF) events, currently missing from existing structured databases. However, this information originates from multiple sources (news articles, government records, and weather records among others) and may cover several topics. Furthermore, these topics may be disproportionately covered on the web. The large size and heterogenous nature of web information render manual review difficult. To address this challenge, we have developed a multi-label text classification model, FF-BERT. FF-BERT is designed to classify FF-related web paragraphs into one or more of seven categories: (1) Damage and Economic Impact (DI), (2) Fatalities, Injuries, and Rescue (FIR), (3) Hydrometeorology (HM), (4) Warning and Emergency (WE), (5) Response and Recovery (RR), (6) Public Health (PH), and (7) Mitigation (MG). To develop FF-BERT, we labeled 21,180 paragraphs from FF-related webpages and performed experiments with multiple model architectures based on the widely used language model Bidirectional Encoder Representation from Transformers (BERT). Our final model outperforms the baseline by 11.83%, as measured by the micro-F1 score. In addition, FF-BERT significantly improves the prediction of minority labels (RR-32.1%, PH-260.4%, and MG-138.6%). We demonstrate using real world examples that FF-BERT can be used to uncover new information about flash flood events. This information can be used to enhance existing databases, such as NOAA’s Storm Events Database.

Recent advances in deep learning have greatly facilitated the automated segmentation of ultrasound images, which is essential for nodule morphological analysis. Nevertheless, most existing methods depend on extensive and precise annotations by domain experts, which are labor-intensive and time-consuming. In this study, we suggest using simple aspect ratio annotations directly from ultrasound clinical diagnoses for automated nodule segmentation. Especially, an asymmetric learning framework is developed by extending the aspect ratio annotations with two types of pseudo labels, i.e., conservative labels and radical labels, to train two asymmetric segmentation networks simultaneously. Subsequently, a conservative-radical-balance strategy (CRBS) strategy is proposed to complementally combine radical and conservative labels. An inconsistency-aware dynamically mixed pseudo-labels supervision (IDMPS) module is introduced to address the challenges of over-segmentation and under-segmentation caused by the two types of labels. To further leverage the spatial prior knowledge provided by clinical annotations, we also present a novel loss function namely the clinical anatomy prior loss. Extensive experiments on two clinically collected ultrasound datasets (thyroid and breast) demonstrate the superior performance of our proposed method, which can achieve comparable and even better performance than fully supervised methods using ground truth annotations.

Deep learning techniques have been investigated for the computer-aided diagnosis of thyroid nodules in ultrasound images. However, most existing thyroid nodule detection methods were simply based on static ultrasound images, which cannot well explore spatial and temporal information following the clinical examination process. In this paper, we propose a novel video-based semi-supervised framework for ultrasound thyroid nodule detection. Especially, considering clinical examinations that need to detect thyroid nodules at the ultrasonic probe positions, we first construct an adjacent frame guided detection backbone network by using adjacent supporting reference frames. To further reduce the labour-intensive thyroid nodule annotation in ultrasound videos, we extend the video-based detection in a semi-supervised manner by using both labeled and unlabeled videos. Based on the detection consistency in sequential neighbouring frames, a pseudo label adaptation strategy is proposed for the refinement of unpredicted frames. The proposed framework is validated on 996 transverse viewed and 1088 longitudinal viewed ultrasound videos. Experimental results demonstrated the superior performance of our proposed method in the ultrasound video-based detection of thyroid nodules.

This paper discusses the implementation of a Deep Reinforcement Learning policy, based on DQN, which optimizes the navigation of the UAV to the front of wind turbine blades. The UAV was trained in simulation using Unreal Engine V4.27 coupled with AirSim. The action space of the UAV was discretized while allowing 6 different actions to be executed. A Yolov5 network trained with images of simulated wind turbines was used for detection and tracking, providing the DQN policy with state information, upon which it has been trained. In addition to this, the dynamic reward has been implemented, which combined both navigation and inspection objectives in the final evaluation of actions. Our tests showed that after 7500 time-steps the exploration rate reached near 0, the mean length of the episodes increased from 10 down to 30, but the mean reward increased from around -60 to stabilizing the output at 26. These results suggest that the proposed method is a promising solution to optimizing the autonomous inspection of wind turbines with UAVs.

Optical methods are becoming powerful clinical tools for non-invasive diagnosis and brain-computer interfacing. The recent development of near infrared spectroscopy, optical coherence tomography, and photoacoustic imaging enable quantitative mapping of brain activity and microscale histological imaging. In this work, we presented principles and applications of these optical methods.