The geographic location information of developers is important for understanding the global distribution of open source activities and formulating regional policies. However, a substantial number of developer accounts on the GitHub platform lack geographic location information, limiting the comprehensive analysis of the geographic distribution of the global open source ecosystem. This study proposed a hierarchical geographic location prediction framework based on multidimensional feature fusion. By integrating three major categories of multidimensional features—temporal behavior, linguistic culture, and network characteristics—the framework established a four-tier progressive prediction mechanism consisting of rule-driven rapid positioning, name cultural inference, time zone cross-validation, and a deep learning ensemble. Experiments conducted on a large-scale dataset built from 50000 globally active developers demonstrated that this method successfully predicted the geographic locations of 82.52% of the developers. Among these, the name cultural inference layer covered most users with an accuracy of 0.7629, whereas the deep learning ensemble layer handled the most complex cases with an accuracy of 0.7557. A comparative analysis with the prediction results from the Moonshot large language model validated the superiority of the proposed method in complex geographic inference tasks.
With the rapid development of open source communities, the number of GitHub projects has increased exponentially. However, a considerable portion of these projects lack explicit topic labels, creating challenges for developers in technology selection and project retrieval processes. Existing topic generation methods rely primarily on supervised learning paradigms that suffer from strong dependencies on high-quality annotated data and other limitations. This study addresses the accuracy and efficiency issues in open source community project topic annotation by conducting the first comprehensive study on the application effectiveness of large language models in GitHub project topic prediction tasks. We constructed a dataset containing 3000 popular GitHub projects that were selected based on a quantitative metric specifically designed to evaluate the activity and influence of open source projects, encompassing multidimensional features including repository names, README documents, and description information. Comparative experiments were conducted using several mainstream large language models from domestic and international sources including Claude 3.7 Sonnet, DeepSeek-V3, Gemini 2.0 Flash, GPT-4o, and Qwen-Plus. The results demonstrated that Claude 3.7 Sonnet achieved optimal performance across most evaluation metrics, and as the dataset scale expanded, the performances of all models tended to stabilize. The experiments proved that large language models exhibited excellent applicability in project topic annotation tasks, although significant performance differences existed among different models. These findings provide an important reference foundation for open source community project management and intelligent annotation system design.
With the development of the digital economy, the explosion of mobile Internet, and the rise of cloud business models, ecological industries have shown significant vitality in the capital market. This article explores the “liquidity” element of the ecological industry, analyzes its value and role in building the ecological industry, and proposes suggestions for building an ecological industry cluster around open-source communities and open-source talents as the core. By formulizing an equation for “open-source index,” for the first time, the article provides a method that can quantitatively identify the importance of the open-source ecosystem for building the industrial ecology. Finally, the article proposes specific strategies and improvement suggestions based on the domestic mainstream ecological industry development cases of open-source ecological construction. For future work, the article identifies several directions, including but not limited to quantified liquidity study of the industrial ecology, cross industry collaboration, policy and legal environment, and security.
Data structures and algorithms (DSA), as a core course in computer science education, play a key role in cultivating programming skills and algorithmic thinking of students. Visualization can significantly enhance teaching effectiveness and deepen student understanding in DSA education. However, existing DSA visualization tools often rely on manually written visualization codes that lead to limitations such as limited coverage, high maintenance costs, and lack of interactivity; hence, the needs of dynamic demonstrations and personalized teaching are difficult to meet. With the outstanding performance of large language models (LLMs) in code generation, automated DSA visualization has become a promising possibility. Therefore, this study proposed an interactive visualization code generation method based on the reasoning and acting (ReAct) AI agent framework, aiming to address the low automation and insufficient interactivity of traditional visualization tools. By leveraging the code generation capabilities of LLMs and integrating with the data structure visualization (DSV) platform interface, the proposed method transformed Python-based DSA code into interactive, executable, and dynamically visualized code, thereby enhancing teaching clarity and learning experience. To systematically evaluate the effectiveness of the method, we constructed a dataset of 150 pairs of DSA code and corresponding DSV visualization code and compared three approaches—direct prompting, chain-of-thought prompting, and the ReAct AI agent approach—across several mainstream LLMs. The experimental results showed that the proposed ReAct AI agent-based method significantly outperformed the other approaches in terms of the compilation rate, execution rate, and usability rate, with the best performance observed in the DeepSeek-R1 model. This demonstrated notable improvements in the accuracy and interactivity of generated visualization code. This research confirms the feasibility and advantages of integrating LLMs with agent frameworks in DSA visualization teaching, offering a novel path toward building efficient, personalized, and automated tools for computer programming education.
This paper introduces ATBench, a benchmark designed for evaluating analysis trajectories in end-to-end data analysis tasks, to address the limitations in granularity and domain coverage present in current benchmarks. Analysis trajectories represent the process in which an agent iteratively poses questions, derives insights, and formulates conclusions around a specific analysis goal via iterative interactions. Leveraging both existing benchmarks and real Kaggle task data, we constructed 151 evaluation datasets spanning eight distinct domains by employing an annotation strategy that balances goal-driven and exploratory approaches. Additionally, we propose a fine-grained evaluation metric, the analysis trajectory score, to assess an agent's coherent analytical capabilities during end-to-end data analysis tasks. Experimental results demonstrate that ATBench exhibits strong stability and discriminative power, effectively distinguishing performance differences among models in analytical tasks. The results also reveal the limitations in agents’ abilities for coherent analysis and insight discovery, thereby providing data-driven support for future improvements.
To address the limited robustness of large multimodal models (LMMs) in complex visual scenarios, such as identifying responsibility for fallen trees, which emanates from their reliance on single-path reasoning. This study proposes a novel reasoning optimization method based on Beam Search Chain-of-Thought (BS-CoT). Conventional models often fall into a “first-impression” trap, in which an initial incorrect inference leads to an irreversible analytical failure. The proposed BS-CoT method counteracts this by exploring and evaluating multiple potential inference paths in parallel. It maintains a diverse set of hypotheses about the scene, continuously pruning less likely hypotheses, which effectively overcomes the tendency to commit to a single, fallacious line of reasoning. This significantly enhances visual decision-making capabilities in complex and noisy environments. To validate its efficacy, we constructed a specialized dataset capturing a wide array of treefall incidents in urban governance. Experimental results demonstrated that the proposed method achieved substantial improvements in both event recall and key information capture rates compared with baseline models. This research not only provides a reliable technical solution for visual decision-making challenges in urban public safety but also introduces a new, more robust paradigm for improving the reasoning reliability of large models in critical applications.
中国综合性科技类核心期刊(北大核心)
