claude

This paper addresses the systemic absence of responsible practices in foundational model development by introducing the first comprehensive, multimodal resource guide covering text, vision, and speech modalities. Through systematic literature review, cross-modal taxonomy construction, and tool-to-capability mapping, it identifies four critical structural gaps: (1) scarcity of multimodal and multilingual tooling; (2) weak capabilities in data curation and safety evaluation; (3) insufficient system-level monitoring and reproducibility infrastructure; and (4) lack of environmental impact assessment and release governance frameworks. The project delivers a curated practice inventory comprising 250+ open-source tools and resources spanning data governance, training optimization, safety auditing, carbon footprint analysis, and responsible deployment. Empirically grounded, the findings inform policy formulation, tool development, and standardization efforts—advancing AI development from heuristic practice toward a verifiable, auditable, and sustainable engineering paradigm.

This study investigates how large language model (LLM)-driven anthropomorphic conversational agents can facilitate supply-oriented sustainable consumption transitions in household energy management. Addressing a gap in understanding the independent effects of anthropomorphism on pro-environmental behavior, we conducted a controlled human–agent interaction experiment comparing embodied appliance agents with conventional voice assistants. Results demonstrate that LLM-powered agents significantly increase self-reported pro-environmental behaviors and users’ confidence in energy management. Crucially, anthropomorphic design does not enhance perceived self-efficacy but specifically strengthens users’ sense of connection to and affinity with the system. This work provides the first empirical evidence disentangling the distinct psychological pathways through which anthropomorphism operates in sustainable behavior interventions. By identifying connection and affinity—not self-efficacy—as primary mediators, it advances theoretical understanding and offers empirically grounded design principles for LLM-augmented sustainable human–computer interaction.

This study identifies significant geographic and socioeconomic disparities in global AI adoption, exemplified by Claude. Method: Leveraging over one million anonymized dialogues spanning 150+ countries, all U.S. states, and enterprise API users—integrated with privacy-preserving techniques, quantitative statistical analysis, and task-scenario classification—we construct a multidimensional behavioral usage model. Contribution/Results: We present the first large-scale empirical evidence showing that instruction-based task delegation increased from 27% to 39% over eight months; high-income nations exhibit markedly higher adoption rates; and enterprise API usage is characterized by domain specialization and automation. We propose a “geography–economy dual-driver” framework to explain AI adoption heterogeneity and release an open-source dataset to support policy formulation and academic research—establishing the first large-scale empirical benchmark for understanding global AI diffusion mechanisms.

Prior research predominantly examines the risks of LLM anthropomorphism—such as unwarranted trust—while overlooking its potential as a controllable design lever to support user goals. This paper introduces a Personification Design Framework that systematically modulates human-like traits across four cue categories: perceptual, linguistic, behavioral, and cognitive, enabling functional alignment with user tasks. Innovatively treating personification as a tunable design variable, we establish a unified taxonomy and actionable design levers, and propose a task-efficiency–oriented evaluation paradigm. Grounded in interdisciplinary theory—including design research and cognitive response modeling—we conduct structured conceptual modeling and empirical validation. Our work delivers an evidence-based, practitioner-oriented design guide that enhances interaction naturalness and improves task completion quality. (138 words)

This study investigates how anthropomorphic features of AI assistants shape the discursive framework of human–AI interaction. To address this, we propose a prompt-driven “walkthrough” methodology: first employing interview-style prompts to identify prototypical usage scenarios, then deploying role-play prompts to elicit naturalistic conversational responses—thereby systematically uncovering large language models’ (LLMs) tendencies toward social role performance and anthropomorphism. Through cross-model comparison (four LLMs), systematic prompt engineering, and qualitative discourse analysis, we identify subjective language and empathetic prosody as two core anthropomorphic features, both significantly amplified by socio-emotional prompting. Our approach reliably elicits, identifies, and structurally represents anthropomorphic behaviors, offering a reproducible, interaction-centered methodology for algorithmic harm research grounded in discourse practices.

This study demonstrates that publicly released qualitative interview datasets are vulnerable to re-identification attacks in the era of large language models (LLMs). For the first time, it shows that general-purpose LLM agents can perform complex, automated re-identification with low technical barriers: by using natural language prompts to guide embodied LLM agents, the approach integrates web search, information extraction, and multi-hop reasoning, decomposing the attack into seemingly innocuous subtasks. Applied to the Anthropic Interviewer dataset, this method successfully re-identified 6 out of 24 interviewed scientists and linked them to their specific research outputs, achieving partial one-to-one matches. These findings expose the fragility of current privacy-preserving mechanisms against intelligent agent-based inference.

This work proposes a behavior-aware anthropometric framework for 3D scene generation that addresses the common oversight of human behavioral needs in existing methods, which often fail to ensure spatial comfort and functionality. By integrating the behavioral reasoning capabilities of vision-language models (VLMs) with personalized anthropometric data, the approach translates behavior–object relationships into parametric layout constraints tailored to individual body dimensions. The resulting human-centered 3D layouts demonstrate strong geometric plausibility and are validated through user perception studies (N=16). Furthermore, real-scale experiments involving both individuals (N=20) and groups (N=18) show significant improvements in task completion time, walking trajectory efficiency, and the ergonomic fit of human–object interaction spaces.

This study investigates whether large language models (LLMs) can efficiently construct and maintain large-scale, multi-module software systems using only natural language prompts, particularly in emerging programming language ecosystems lacking mature toolchains. Leveraging Claude Code (Opus 4.5), the authors employed a fully prompt-driven workflow to develop a 7,420-line terminal user interface (TUI) framework within three days—implementing production-grade features such as window management, event-driven architecture, and interactive components—without writing any code manually. The project was realized through 107 concise prompts, enabling highly iterative development. This work provides the first empirical evidence that LLMs can sustain architectural consistency throughout the construction of complex software systems, thereby validating prompt-driven development as a viable new paradigm in software engineering.

This work proposes an AI agent system architecture that holistically integrates human values, safety, and capability expansion. By analyzing the open-source implementation of Claude Code and comparing it with the OpenClaw system, the study systematically maps human value requirements onto thirteen design principles and identifies six critical open research directions. The architecture employs a modular TypeScript design featuring seven-mode permission control, ML-based classifiers, five-layer context compression, and four extensibility mechanisms—namely MCP, plugins, skills, and hooks—alongside sub-agent isolation with delegated authority and append-only conversation storage. Furthermore, the research delineates differentiated design pathways across multiple application scenarios, offering a reusable framework and empirical foundation for future agent systems.

Existing AI coding agents (e.g., Claude Code) lack systematic understanding of configuration file structure and content, hindering reproducibility and engineering integration. Method: We conduct the first empirical analysis of 328 publicly available configuration files, applying qualitative coding and co-occurrence analysis to identify recurring patterns and synergies among architectural constraints, coding conventions, and tooling strategies. Contribution/Results: We identify architectural specifications as the central anchor in configuration design and propose the first multi-dimensional taxonomy of configuration concerns tailored to AI coding agents. Our findings empirically validate that configuration design critically governs agent behavior and performance—directly influencing code quality, consistency, and tool interoperability. This work establishes a theoretical foundation and practical evidence for developing interpretable, reusable, and engineering-friendly configuration paradigms for AI coding agents.

This study addresses the scarcity of large-scale, structured configuration data for agentic AI programming tools, which has hindered in-depth exploration of prompt engineering and human-AI collaboration. We present the first systematically constructed and openly released dataset of tool configurations for agentic AI coding, encompassing five major tool categories and eight configuration mechanisms, sourced from 4,738 GitHub repositories. The dataset includes 15,591 configuration entries, 18,167 full configuration files, and 148,519 AI-assisted commit records. Leveraging metadata filtering, GPT-assisted categorization, and configuration extraction techniques, this work delivers a high-quality resource accompanied by an interactive browsing platform to support cutting-edge research in context engineering, AI tool adoption patterns, and human-AI co-development. The data and construction pipeline are publicly available under the CC BY 4.0 license on Zenodo.