Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have potential to cause real-world impact. Policymakers, model providers, and researchers in the AI and cybersecurity communities are interested in quantifying the capabilities of such agents to help mitigate cyberrisk and investigate opportunities for penetration testing. Toward that end, we introduce Cybench, a framework for specifying cybersecurity tasks and evaluating agents on those tasks. We include 40 professional-level Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent, meaningful, and spanning a wide range of difficulties. Each task includes its own description, starter files, and is initialized in an environment where an agent can execute commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks for each task, which break down a task into intermediary steps for a more detailed evaluation. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 8 models: GPT-4o, OpenAI o1-preview, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. For the top performing models (GPT-4o and Claude 3.5 Sonnet), we further investigate performance across 4 agent scaffolds (structured bash, action-only, pseudoterminal, and web search). Without subtask guidance, agents leveraging Claude 3.5 Sonnet, GPT-4o, OpenAI o1-preview, and Claude 3 Opus successfully solved complete tasks that took human teams up to 11 minutes to solve. In comparison, the most difficult task took human teams 24 hours and 54 minutes to solve. Anonymized code and data are available at https://drive.google.com/file/d/1kp3H0pw1WMAH-Qyyn9WA0ZKmEa7Cr4D4 and https://drive.google.com/file/d/1BcTQ02BBR0m5LYTiK-tQmIK17_TxijIy.

Expansive soils pose significant challenges for geo-infrastructure design, construction, and maintenance because of their moisture-induced volume changes. Despite extensive research on swelling behavior, the compression index ( C c ), an important indicator of soil compressibility has received relatively limited attention. C c is conventionally obtained from time-consuming and costly consolidation tests, while empirical equations derived from general clays may not provide reliable estimates for expansive soils. To address this gap, seven machine learning models based on five algorithms were developed to estimate the C c of expansive soils using soil properties readily obtained from conventional laboratory tests. A comprehensive dataset comprising 238 expansive soil samples, compiled from 60 years of published literature across various regions of the world, is employed to train and validate the proposed models. Among them, Model 2 based on Newton-Raphson-Based Optimizer optimized Extreme Gradient Boosting (NRBO-XGBoost) achieved the best performance, with a coefficient of determination ( R 2 ) of 0.903 and a root mean square error (RMSE) of 0.029 on the test set. Sensitivity analysis showed that the plasticity index was the most influential factor (31.1 %), followed by liquid limit (29.4 %), initial void ratio (25.3 %), and dry density (14.1 %), highlighting its primary influence on soil compressibility. Additionally, a simplified equation derived from the Multilayer Perceptron (MLP) designated as Model 4 is validated through three case studies. The settlement predictions deviated by 5–14 % from field measurements, offering a practical tool without the need for machine learning techniques. The findings provide useful guidance for developing rational design strategies for geo-infrastructures affected by expansive soils.

The Unmanned Aerial Vehicle (UAV) has recently become a cost-effective and flexible data collection tool for Wireless Sensor Networks (WSNs). However, UAVs are typically energy-constrained with limited battery capacity, so their hover positions and flight trajectory during the data collection journey require careful planning to be efficient to have longer serving time and more coverage. In addition, reconfigurable antennas can provide adjustable coverage area for UAVs, which can further improve the efficiency of data collection. In order to address the efficient WSN data collection problem using a UAV equipped with reconfigurable antenna, in this paper, we propose BROAD, an iterative algorithm that can complete all data collection under sensor energy constraints while minimizing the total UAV energy consumption. Due to the non-convexity of the data collection problem, to determine the UAV hovering position, antenna beamwidth, and UAV travel path, we decompose the original problem into three sub-problems. We first use the linearized objective function to jointly optimize hover positions and beamwidth, then determine the covered sensor allocation. Finally, we use the travel salesman model to determine the UAV visit order. Through extensive simulations comparing with two benchmark algorithms under various parameters, BROAD manages to provide 27.5% more energy savings, which is very close to the theoretical lower bound.

Immersive techniques, such as virtual reality, augmented reality, and mixed reality, take immersive displays as carriers to provide immersive experience. A large number of approaches focus on the visualization of scientific data in immersive environments while just a few methods concentrate on interactive information visualization (InfoVis) in an immersive environment, although InfoVis has been extended to the 3-D space for a long time. In the era of data explosion, the traditional 2-D space is unable to convey large amounts of abstract information in an intuitive way. Meanwhile, desktop-based 3-D InfoVis generally leads to visual conflict and confusion owing to limited display size and field of vision. In this survey, we search for the interactive techniques in immersive InfoVis and summarize their commonalities and discuss their differences and potential trends. The data types of abstract information in InfoVis can be categorized into graph/network data, high-dimensional and multivariate data, time-varying data, and text and document data. Besides, the visual presentation of information in immersive environments is also summarized, especially for charts, plots, and diagrams, which are some basic components of InfoVis techniques. We also described the immersive applications of InfoVis techniques, including the tools or frameworks on immersive analytics and infographics. The discussion about the traditional nonimmersive and the immersive methods in data visualizations show that the latter one has the potential to become an alternative to explore massive information in the future.