本文提出HIAR模型，用于处理不规则时间序列的四成分观测数据，并通过Kalman滤波器估计参数，应用于Sentinel-2图像分析森林边缘。
This study addresses the integration barriers hindering the widespread adoption of open-source Self-Sovereign Identity (SSI) frameworks, which stem from complex toolchains, insufficient abstraction, and outdated documentation. Through a developer experience study, nine developers with decentralized identity backgrounds were tasked with completing credential lifecycle operations using Walt.id, Traction, and MetaMask. Combining task difficulty ratings with qualitative thematic analysis, the research systematically identifies structural abstraction gaps within the SSI ecosystem for the first time. Findings reveal that passive operations—such as credential reception—are relatively mature, whereas active construction tasks—like schema customization—exhibit significant architectural friction. To mitigate these challenges, the paper proposes three actionable improvements: web-based sandbox environments, AI-assisted schema generators, and executable documentation, offering practical pathways to lower the barrier to SSI integration.
This study addresses the persistent theory–practice gaps in agile software development—specifically theoretical, temporal, and translational disconnects—by convening the Second Workshop on Agile Practice and Research, which brought together academic and industry stakeholders. Through qualitative methods including structured small-group collaboration, reflective discussions, and synthesis of outputs, the work investigates the root causes of these gaps. Building on these insights, it proposes four propositions to enhance intersection mechanisms and distills them into three research imperatives centered on open science, theoretical rigor, and value orientation. The study concludes with four practical recommendations: improving scholarly communication, aligning with industry needs, incentivizing sustained collaboration, and integrating educational practices—collectively advancing agile research toward greater transparency, methodological rigor, and practical relevance.
This study addresses the challenge of reducing computational costs and resource barriers for large language models in Brazilian Portuguese question answering. We systematically evaluate parameter-efficient fine-tuning (PEFT) methods—including LoRA, DoRA, QLoRA, and QDoRA—on BERTimbau Base and Large, and compare their efficiency and effectiveness against generative large language models such as Tucano and Sabiá, using the SQuAD-BR dataset. Our results show that LoRA achieves 95.8% of the baseline performance on BERTimbau-Large with 73.5% less training time, while higher learning rates yield up to a +19.71 F1 improvement. Although generative models remain competitive, they require 4.2× more GPU memory and 3× longer training times. The findings highlight the efficacy of PEFT approaches and their potential for enabling green AI in low-resource language settings.
This work addresses the common gap in students’ practical experience with user interaction in agile development and their limited understanding of the capabilities and limitations of generative AI in requirements engineering. To bridge this gap, the study introduces an innovative approach that employs a generative AI–powered virtual stakeholder simulation, guided by meta-prompting to facilitate student-led requirement interviews. The method integrates agile practices such as user story mapping and impact mapping for requirements elicitation and documentation, followed by structured reflective discussions to deepen students’ awareness of the technical boundaries and ethical implications of AI tools. Designed to be model-agnostic, the approach demonstrates flexibility and reusability across contexts. Multi-semester teaching evaluations confirm its effectiveness in enhancing students’ integrated competencies in cutting-edge agile requirements engineering and the synergistic application of generative AI.
本文提出HIAR模型，用于处理不规则时间序列的四成分观测数据，并通过Kalman滤波器估计参数，应用于Sentinel-2图像分析森林边缘。
This study addresses the integration barriers hindering the widespread adoption of open-source Self-Sovereign Identity (SSI) frameworks, which stem from complex toolchains, insufficient abstraction, and outdated documentation. Through a developer experience study, nine developers with decentralized identity backgrounds were tasked with completing credential lifecycle operations using Walt.id, Traction, and MetaMask. Combining task difficulty ratings with qualitative thematic analysis, the research systematically identifies structural abstraction gaps within the SSI ecosystem for the first time. Findings reveal that passive operations—such as credential reception—are relatively mature, whereas active construction tasks—like schema customization—exhibit significant architectural friction. To mitigate these challenges, the paper proposes three actionable improvements: web-based sandbox environments, AI-assisted schema generators, and executable documentation, offering practical pathways to lower the barrier to SSI integration.
This study addresses the persistent theory–practice gaps in agile software development—specifically theoretical, temporal, and translational disconnects—by convening the Second Workshop on Agile Practice and Research, which brought together academic and industry stakeholders. Through qualitative methods including structured small-group collaboration, reflective discussions, and synthesis of outputs, the work investigates the root causes of these gaps. Building on these insights, it proposes four propositions to enhance intersection mechanisms and distills them into three research imperatives centered on open science, theoretical rigor, and value orientation. The study concludes with four practical recommendations: improving scholarly communication, aligning with industry needs, incentivizing sustained collaboration, and integrating educational practices—collectively advancing agile research toward greater transparency, methodological rigor, and practical relevance.
This study addresses the challenge of reducing computational costs and resource barriers for large language models in Brazilian Portuguese question answering. We systematically evaluate parameter-efficient fine-tuning (PEFT) methods—including LoRA, DoRA, QLoRA, and QDoRA—on BERTimbau Base and Large, and compare their efficiency and effectiveness against generative large language models such as Tucano and Sabiá, using the SQuAD-BR dataset. Our results show that LoRA achieves 95.8% of the baseline performance on BERTimbau-Large with 73.5% less training time, while higher learning rates yield up to a +19.71 F1 improvement. Although generative models remain competitive, they require 4.2× more GPU memory and 3× longer training times. The findings highlight the efficacy of PEFT approaches and their potential for enabling green AI in low-resource language settings.
This work addresses the common gap in students’ practical experience with user interaction in agile development and their limited understanding of the capabilities and limitations of generative AI in requirements engineering. To bridge this gap, the study introduces an innovative approach that employs a generative AI–powered virtual stakeholder simulation, guided by meta-prompting to facilitate student-led requirement interviews. The method integrates agile practices such as user story mapping and impact mapping for requirements elicitation and documentation, followed by structured reflective discussions to deepen students’ awareness of the technical boundaries and ethical implications of AI tools. Designed to be model-agnostic, the approach demonstrates flexibility and reusability across contexts. Multi-semester teaching evaluations confirm its effectiveness in enhancing students’ integrated competencies in cutting-edge agile requirements engineering and the synergistic application of generative AI.