Agentic coding tools receive goals written in natural language, break them down into specific tasks, and write or execute code with minimal human intervention. Central to this process are agent context files (e.g., AGENTS.md and CLAUDE.md) that provide persistent, project-level instructions. In this article, we conduct the first large-scale empirical study of 2,303 agent context files from 1,925 repositories to characterize their structure, maintenance, and content. We find that these files are not static documentation but complex, difficult-to-read artifacts that evolve like configuration code through frequent, small additions. Our content analysis of 16 instruction types shows that developers prioritize functional context, such as test procedures (75.9%), implementation details (70.8%), and architecture (68.1%). We also identify a significant gap: non-functional requirements such as security (14.8%) and performance (14.5%) are rarely specified. These findings indicate that while developers use context files to make agents functional, they provide few guardrails to ensure that agent-written code is secure or performant, highlighting the need for improved tools and practices.
Deploying neural network models to edge devices is becoming increasingly popular because such deployment decreases the response time and ensures better data privacy of services. However, running large models on edge devices poses challenges because of limited computing resources and storage space. Researchers have therefore proposed various model compression methods to reduce the model size. To balance the trade-off between model size and accuracy, conventional model compression methods require manual effort to find the optimal configuration that reduces the model size without significant degradation of accuracy. In this article, we propose a method to automatically find the optimal configurations for quantization. The proposed method suggests multiple compression configurations that produce models with different size and accuracy, from which users can select the configurations that suit their use cases. Additionally, we propose a retraining method that does not require any labeled datasets for retraining. We evaluated the proposed method using various neural network models for classification, regression and semantic similarity tasks, and demonstrated that the proposed method reduced the size of models by at least 30% while maintaining less than 1% loss of accuracy.We compared the proposed method with state-of-the-art automated compression methods, and showed that it can provide better compression configurations than existing methods.
Snoring is one of the most important health problems. It could be a sign of sleep apnea which leads to a potentially serious sleep disorder. Thus, detecting snoring sounds is the first step to diagnosing sleep disorders. Due to improving edge device performance, we can use edge devices, or mobile phones, to detect snoring during sleep. Edge devices allow us to deploy a machine learning model, which is a widely used technique to detect snoring. However, snoring sounds are privacy-sensitive data because they contain activity information through participants’ sleep. This paper aims to detect snoring sounds on edge devices while preserving data privacy. We designed and implemented a convolutional-based neural network model that utilizes fewer commuting resources while maintaining the model’s accuracy. The proposed model contains six two-dimensional convolutional and two fully connected layers. As a result, the accuracy is improved by 1.96% and the model size is reduced by 66%. Due to the smaller model size, the proposed model is able to maintain the model accuracy with lower required communication cost for federated learning when compared to the original accuracy.
As the number of edge devices rapidly multiplies into the millions in the post-5G era, there is a simultaneous surge in demand for job execution driven by the data generated by these devices. Ultra-distributed computing systems have emerged to support this notable proliferation of interconnected devices and in addressing the unprecedented data generated over low-latency networks. To ensure the high availability and reliability of these systems, an efficient job scheduling method is required to schedule jobs across available resources. However, there are currently no large-scale benchmarks available to evaluate job scheduling methods. This paper introduces large-scale benchmarks of up to one million devices for job shop scheduling problems. We evaluate the makespan reduction of widely used combinatorial optimizations on these benchmarks, including simulated annealing, ant colony, tree Parzen search, particle swarm, artificial bee colony, cuckoo search, whale, grey wolf, firefly, and bat optimizations. We investigate the impact of execution time on finding the optimal makespan of job shop scheduling for each method. The experimental results can be used to guide the selection of a job scheduling method for particular applications.
Currently, large language models (LLMs) have been utilized in many aspects of natural language processing. However, due to their significant size and high computational demands, large computational resources are required for deployment. In this research, we focus on the automated approach for size reduction of such a model. We propose the framework to perform the automated pruning based on combinatorial optimization. Two techniques were particularly studied, i.e., particle swarm optimization (PSO) and whale optimization algorithm (WOA). The model pruning problem was modeled as a combinatorial optimization task whose the goal is to minimize model size while maintaining model accuracy. The framework systematically explores the search space to identify the most optimal pruning configurations, removing redundant or non-contributory parameters. The two optimizations, PSO and WOA, were evaluated for their ability to efficiently navigate the search space. As a result, with PSO, the proposed framework can reduce the model size of Llama-3.1-70B by 13.44% while keeping the loss of model accuracy at 19.25%; with WOA, the model size reduction is 12.07% with 22.81% loss of model accuracy. Since accuracy degradation may occur during pruning process, the framework integrates the post-process to recover the model accuracy. After this process, the pruned model loss can reduce to 12.72% and 14.83% using PSO and WOA, respectively.