dependency analysis

This work addresses the challenge of implicit, fragmented, and recursive dependencies in large language model (LLM) development, which are difficult to trace manually and lead to licensing non-compliance, evaluation bias, and documentation inconsistencies. The authors propose ModSleuth, a system that formalizes LLM dependency types for the first time and employs an operation-centric modeling approach. ModSleuth leverages agents to automatically and recursively extract and verify dependencies from public artifacts, accurately distinguishing between direct and indirect dependencies while resolving entity alignment across disparate names, versions, and repositories. Evaluated on four LLM releases, the system identifies 1,060 source-verified dependencies, uncovering critical issues such as multi-hop licensing obligations, training-evaluation coupling, and documentation mismatches. The system and its dependency graphs are publicly released.

Pre-trained models (PTMs) are increasingly adopted as novel software dependencies—termed “Software Dependencies 2.0”—yet their reuse and integration practices in open-source projects, along with associated maintainability and reliability implications, remain poorly understood. Method: We conduct the first systematic empirical study using a mixed-methods approach: statistically significant sampling of 401 GitHub repositories from the PeaTMOSS dataset, combined with quantitative pattern mining and in-depth qualitative case analysis. Contribution/Results: We find widespread deficiencies in PTM versioning, documentation, and dependency tracking. We identify three canonical PTM reuse pipelines and uncover complex, cross-stage, asymmetric inter-model dependency structures. Critically, we empirically define and characterize the first structured usage paradigm for Software Dependencies 2.0, establishing foundational theory and evidence to guide model-aware software engineering practices and tooling.

In hardware development, 3D CAD models exhibit highly complex dependency structures—often comprising thousands of components—leading to significant challenges in impact analysis, cross-role collaboration, and synchronization. This complexity exposes nine critical issues spanning traceability, navigability, and consistency. To address this gap, we conducted a thematic analysis of 100 online forum discussions and semi-structured interviews with 10 senior hardware designers, systematically identifying and categorizing these pain points for the first time. Building on these findings, we propose the “dependency-aware collaboration” design paradigm and introduce a corresponding framework featuring dependency visualization, change-propagation alerts, and contextual synchronization for collaborative editing, guided by six design principles. Our work fills a theoretical void in CSCW research on hardware co-design and provides empirically grounded, actionable guidelines for next-generation CAD collaboration tools.

This study investigates whether large language model (LLM) agents encode the dependency structure among tool calls within their internal representations during execution. By applying low-capacity edge probing to the residual stream of Qwen3-32B and integrating counterfactual perturbations—distinguishing between value corruption and structural disruption—with activation patching and a multi-hop interaction benchmark, this work presents the first structural probing of LLM tool-call dependency graphs. Results demonstrate that the model linearly decodes a non-positional, propagative directed dependency graph, significantly outperforming random and positional baselines. Notably, this signal vanishes in single-step planning settings, confirming its specific association with multi-hop reasoning and revealing the LLM’s intrinsic capacity to represent abstract task topologies.

In microservice architectures, implicit semantic dependencies—inter-service couplings arising from highly similar business logic despite the absence of explicit invocations or data sharing—pose significant change consistency risks; conventional dependency analysis techniques fail to detect them. This paper proposes the “Semantic Dependency Matrix,” the first automated detection framework targeting logical semantics. It integrates code semantic parsing, runtime behavior modeling, and fine-grained similarity measurement to extract and quantify semantic coupling directly from service implementations. Evaluated on real-world industrial systems, our approach uncovers critical semantic dependencies entirely missed by traditional interface or call-graph analyses. It improves change impact prediction accuracy by 37.2%, thereby substantially enhancing system maintainability and evolutionary robustness.

Current LLM agents lack explicit dependency management for their skills, leading to redundant dependencies, environmental inconsistencies, and security risks. This work proposes the Agent Skill Supply Chain (ASSC) framework, which introduces a systematic model of hybrid dependencies among skills, packages, and services. Inspired by Software Bill of Materials (SBOM), we design SkillDepAnalyzer—a tool that automatically extracts dependency evidence from natural language descriptions to construct skill dependency graphs. Experiments on the SKILL-DEP benchmark demonstrate that our approach significantly outperforms both LLM-based baselines and conventional SBOM tools. An analysis of 1.43 million skills reveals four distinct dependency patterns and enables effective identification and reporting of malicious skills, offering a novel paradigm for secure governance in skill ecosystems.

This study systematically identifies and differentiates functional dependency, overreliance, and addiction-like behaviors in the use of large language models (LLMs) within software engineering. Through a survey of 119 software practitioners, complemented by descriptive statistics and thematic analysis of open-ended responses, the research reveals that LLMs have become deeply integrated into development workflows. Most developers exhibit functional dependency, leveraging LLMs as productive tools without impairment. Overreliance manifests as a preference for consulting LLMs over official documentation or colleagues, potentially compromising solution quality. Addiction-like behaviors are relatively rare but characterized by difficulties in usage moderation and emotional attachment. The findings provide an empirical foundation and a behavioral classification framework for understanding the nuanced impacts of LLM adoption in software development.

This work proposes a paradigm shift from validity-driven to value-driven data dependency discovery, addressing the limitations of traditional approaches that focus narrowly on statistical strength or validity while overlooking holistic value in real-world data governance tasks—such as relevance, redundancy, and lifecycle costs. Drawing on decision theory, the paper formally defines the utility and net value of dependencies and introduces a comprehensive, value-aware framework spanning discovery, validation, selection, and maintenance. Its key innovation lies in unifying task-specific loss reduction and lifecycle cost within a single evaluation framework, integrating budget-constrained optimization, loss–cost learning, and dynamic monitoring mechanisms. This establishes both the theoretical foundation and key technical pathways for value-driven dependency discovery, offering a new direction toward efficient and cost-effective data governance.

This study addresses a structural misalignment between producers and consumers of pretrained language models (PTLMs) on platforms like Hugging Face, which manifests as mismatches in model discovery, documentation, lineage tracing, and governance. Through surveys and qualitative analysis involving 50 model producers and 95 GitHub-based consumers, this work reveals significant discrepancies in how the two groups perceive the placement of critical metadata, motivations for lineage tracking, and priorities in model governance. These findings provide empirical grounding for improving model documentation standards, lineage-tracking tools, and governance frameworks, offering a novel perspective on optimizing the PTLM reuse ecosystem.

This work addresses the challenge that existing code generation methods struggle to effectively model the multi-level dependencies among code entities, often yielding outputs with incomplete logic or inconsistent structure. To overcome this limitation, the paper proposes a graph-constrained code generation framework that explicitly constructs a code dependency graph and leverages it as a structural constraint during generation to ensure both semantic coherence and syntactic consistency. The key innovation lies in decomposing dependency relations into a complementary representation comprising a quantized matrix and sparse low-rank factors, augmented with sparse triplet encodings to capture strong dependencies. This design balances expressive power, memory efficiency, and scalability. Experimental results demonstrate that the proposed approach significantly outperforms current state-of-the-art methods in terms of semantic alignment and structural fidelity.