arXiv subjects
Publications and source records attributed to Qinghua Lu.
Incidents show that AI safety failures often arise across multiple layers. We present a safety-by-design assurance architecture combining model-level supervision, such as Scientist AI, with system-level controls over scaffolds and harnesses, independent verification, monitoring, and evidence infrastructure, supported by governance for accountability and evidence interoperability.
Agent skills externalise reusable agent-facing behavioural knowledge and guidance as persistent artefacts that can be discovered, activated, and interpreted by LLM agents. Although a skill artefact is static at rest, its architectural responsibilities arise in use, when the artefact is selected for a run, bound to context and authority constraints, interpreted by a stochastic agent, and recorded as run evidence. We call this run-specific relation skill-in-use. This paper studies agent skill harnessing: the architectural responsibilities that govern the transition from skill artefacts to skill-in-use, bound the executable consequences associated with skill-in-use, and capture evidence for attribution, verification, repair, and evolution. This paper provides a catalogue of ten empirically grounded architectural patterns (five core, five supporting) for skill harnessing and synthesises them into a reference architecture with four responsibility layers: Supply Chain, Mediation, Execution Control, and Evidence & Feedback. We evaluate the architecture through cross-instantiation across 8 selected systems. The resulting patterns and reference architecture provide a vocabulary and diagnostic frame for analysing skill-harnessing responsibilities across agent systems.
The rise of agentic AI is reshaping software engineering in two intertwined directions: agents are increasingly applied to support software engineering tasks, and Agentic AI systems themselves are complex systems that require re-thinking currently established software engineering practices. To chart a coherent research agenda covering the two directions, we organized the A2SE seminar in Rio de Janeiro, bringing together 18 experts from academia and industry. Through structured presentations, collaborative topic clustering, and focused group discussions, participants identified six thematic areas: Governance, Software Engineering for Agents, Agents for Software Architecture, Quality and Evaluation, Sustainability, and Code, and they prioritized short-term and long-term research directions for each. This paper presents the resulting community-driven, opinionated research agenda, offering the SE community a structured foundation for coordinating efforts at this critical juncture.
Uncertainty in large language model (LLM)-based systems is often studied at the level of a single model output, yet deployed LLM applications are compound systems in which uncertainty is transformed and reused across model internals, workflow stages, component boundaries, persistent state, and human or organisational processes. Without principled treatment of how uncertainty is carried and reused across these boundaries, early errors can propagate and compound in ways that are difficult to detect and govern. This paper develops a systems-level account of uncertainty propagation. It introduces a conceptual framing for characterising propagated uncertainty signals, presents a structured taxonomy spanning intra-model (P1), system-level (P2), and socio-technical (P3) propagation mechanisms, synthesises cross-cutting engineering insights, and identifies five open research challenges.
Mobile apps are essential in daily life but frequently employ deceptive patterns, such as visual emphasis or linguistic nudging, to manipulate user behavior. Existing research largely relies on manual detection, which is time-consuming and cannot keep pace with rapidly evolving apps. Although recent work has explored automated approaches, these methods are limited to intra-page patterns, depend on manual app exploration, and lack flexibility. To address these limitations, we present AppRay, a system that integrates task-oriented app exploration with automated deceptive pattern detection to reduce manual effort, expand detection coverage, and improve performance. AppRay operates in two stages. First, it combines large language model-guided task-oriented exploration with random exploration to capture diverse user interface (UI) states. Second, it detects both intra-page and inter-page deceptive patterns using a contrastive learning-based multi-label classifier augmented with a rule-based refiner for context-aware detection. We contribute two datasets, AppRay-Tainted-UIs and AppRay-Benign-UIs, comprising 2,185 deceptive pattern instances, including 149 intra-page cases, spanning 16 types across 876 deceptive and 871 benign UIs, while preserving UI relationships. Experimental results show that AppRay achieves macro/micro averaged precision of 0.92/0.85, recall of 0.86/0.88, and F1 scores of 0.89/0.85, yielding 27.14% to 1200% improvements over prior methods and enabling effective detection of previously unexplored deceptive patterns.
We present the architectural design and prototype implementation of QUT (Quantum Unit Testing), a framework for unit testing of quantum subroutines. The framework is developed with a focus on usability and simplicity, making the complex theoretical concepts behind quantum unit testing accessible to a wide range of users with diverse backgrounds. This is achieved through the implementation of polymorphic probabilistic assertions, whose evaluation methods adapt to the data types of the arguments used in assertion statements, which may vary according to the context-dependent semantics of quantum subroutines. These arguments can be represented as qubit measurement outcomes, density matrices, or Choi matrices. For each type, the architecture integrates a specific testing protocol - such as quantum process tomography, quantum state tomography, or Pearson's chi-squared test - while remaining flexible enough to incorporate additional protocols in the future. The framework is built on the Qiskit software stack, providing compatibility with a broad range of quantum hardware backends and simulation platforms. Drawing on the reasoning provided by the denotational semantics of quantum subroutines, this work also highlights the key distinctions between quantum unit testing and its classical counterpart.
Safety cases, structured arguments that a system is acceptably safe, are becoming central to the governance of AI systems. Yet, traditional safety-case practices from aviation or nuclear engineering rely on well-specified system boundaries, stable architectures, and known failure modes. Modern AI systems, such as generative and agentic AI, are the opposite. Their capabilities emerge unpredictably from low-level training objectives, their behaviour varies with prompts, and their risk profiles shift through fine-tuning, scaffolding, or deployment context. This study examines how safety cases are currently constructed for AI systems and why classical approaches fail to capture these dynamics. This study introduces comprehensive taxonomies for AI-specific claim types (assertion-based, constraint-based, capability-based), argument types (demonstrative, comparative, causal/explanatory, risk-based, and normative), and evidence families (empirical, mechanistic, comparative, expert-driven, formal methods, operational/field data, and model-based). It then proposes a reusable safety-case template, each of which follows a predefined structure of claims, arguments, and evidence tailored for AI systems. Each template is illustrated by end-to-end patterns that address distinctive challenges, such as evaluation without ground truth, dynamic model updates, and threshold-based risk decisions. The result is a systematic, composable, and reusable approach to constructing and maintaining safety cases that are credible, auditable, and adaptive to the evolving behaviour of generative and frontier AI systems.
Coding standards are essential for maintaining consistent and high-quality code across teams and projects. Linters help developers enforce these standards by detecting code violations. However, manual linter configuration is complex and expertise-intensive, and the diversity and evolution of programming languages, coding standards, and linters lead to repetitive and maintenance-intensive configuration work. To reduce manual effort, we propose LintCFG, a domain-specific language (DSL)-driven, LLM-based compilation approach to automate linter configuration generation for coding standards, independent of programming languages, coding standards, and linters. Inspired by compiler design, we first design a DSL to express coding rules in a tool-agnostic, structured, readable, and precise manner. Then, we build linter configurations into DSL configuration instructions. For a given natural language coding standard, the compilation process parses it into DSL coding standards, matches them with the DSL configuration instructions to set configuration names, option names and values, verifies consistency between the standards and configurations, and finally generates linter-specific configurations. Experiments with Checkstyle for Java coding standard show that our approach achieves over 90% precision and recall in DSL representation, with accuracy, precision, recall, and F1-scores close to 70% (with some exceeding 70%) in fine-grained linter configuration generation. Notably, our approach outperforms baselines by over 100% in precision. A user study further shows that our approach improves developers' efficiency in configuring linters for coding standards. Finally, we demonstrate the generality of the approach by generating ESLint configurations for JavaScript coding standards, showcasing its broad applicability across other programming languages, coding standards, and linters.
Environmental, Social, and Governance (ESG) metric knowledge is inherently structured, connecting industries, reporting frameworks, metric categories, metrics, and calculation models through compositional dependencies, yet in practice this structure remains embedded implicitly in regulatory documents such as SASB, TCFD, and IFRS S2 and rarely exists as an explicit, governed, or machine-actionable artefact. Existing ESG ontologies define formal schemas but do not address scalable population and governance from authoritative regulatory sources, while unconstrained large language model (LLM) extraction frequently produces semantically incorrect entities, hallucinated relationships, and structurally invalid graphs. OntoMetric is an ontology-guided framework for the automated construction and governance of ESG metric knowledge graphs from regulatory documents that operationalises the ESG Metric Knowledge Graph (ESGMKG) ontology as a first-class constraint embedded directly into the extraction and population process. The framework integrates structure-aware segmentation, ontology-constrained LLM extraction enriched with semantic fields and deterministic identifiers, and two-phase validation combining semantic type verification with rule-based schema checking, while preserving segment-level and page-level provenance to ensure traceability to regulatory source text. Evaluation on five ESG regulatory standards shows that ontology-guided extraction achieves 65-90 percent semantic accuracy and over 80 percent schema compliance, compared with 3-10 percent for unconstrained baseline extraction, and yields stable cost efficiency with a cost per validated entity of 0.01-0.02 USD and a 48 times efficiency improvement over baseline.
The rapid rise of autonomous AI systems and advancements in agent capabilities are introducing new risks due to reduced oversight of real-world interactions. Yet agent testing remains nascent and is still a developing science. As AI agents begin to be deployed globally, it is important that they handle different languages and cultures accurately and securely. To address this, participants from The International Network for Advanced AI Measurement, Evaluation and Science, including representatives from Singapore, Japan, Australia, Canada, the European Commission, France, Kenya, South Korea, and the United Kingdom have come together to align approaches to agentic evaluations. This is the third exercise, building on insights from two earlier joint testing exercises conducted by the Network in November 2024 and February 2025. The objective is to further refine best practices for testing advanced AI systems. The exercise was split into two strands: (1) common risks, including leakage of sensitive information and fraud, led by Singapore AISI; and (2) cybersecurity, led by UK AISI. A mix of open and closed-weight models were evaluated against tasks from various public agentic benchmarks. Given the nascency of agentic testing, our primary focus was on understanding methodological issues in conducting such tests, rather than examining test results or model capabilities. This collaboration marks an important step forward as participants work together to advance the science of agentic evaluations.
As frontier AI models are deployed globally, it is essential that their behaviour remains safe and reliable across diverse linguistic and cultural contexts. To examine how current model safeguards hold up in such settings, participants from the International Network for Advanced AI Measurement, Evaluation and Science, including representatives from Singapore, Japan, Australia, Canada, the EU, France, Kenya, South Korea and the UK conducted a joint multilingual evaluation exercise. Led by Singapore AISI, two open-weight models were tested across ten languages spanning high and low resourced groups: Cantonese English, Farsi, French, Japanese, Korean, Kiswahili, Malay, Mandarin Chinese and Telugu. Over 6,000 newly translated prompts were evaluated across five harm categories (privacy, non-violent crime, violent crime, intellectual property and jailbreak robustness), using both LLM-as-a-judge and human annotation. The exercise shows how safety behaviours can vary across languages. These include differences in safeguard robustness across languages and harm types and variation in evaluator reliability (LLM-as-judge vs. human review). Further, it also generated methodological insights for improving multilingual safety evaluations, such as the need for culturally contextualised translations, stress-tested evaluator prompts and clearer human annotation guidelines. This work represents an initial step toward a shared framework for multilingual safety testing of advanced AI systems and calls for continued collaboration with the wider research community and industry.
The rapid emergence of pretrained models (PTMs) has attracted significant attention from both Deep Learning (DL) researchers and downstream application developers. However, selecting appropriate PTMs remains challenging because existing methods typically rely on keyword-based searches in which the keywords are often derived directly from function descriptions. This often fails to fully capture user intent and makes it difficult to identify suitable models when developers also consider factors such as bias mitigation, hardware requirements, or license compliance. To address the limitations of keyword-based model search, we propose PTMPicker to accurately identify suitable PTMs. We first define a structured template composed of common and essential attributes for PTMs and then PTMPicker represents both candidate models and user-intended features (i.e., model search requests) in this unified format. To determine whether candidate models satisfy user requirements, it computes embedding similarities for function-related attributes and uses well-crafted prompts to evaluate special constraints such as license compliance and hardware requirements. We scraped a total of 543,949 pretrained models from Hugging Face to prepare valid candidates for selection. PTMPicker then represented them in the predefined structured format by extracting their associated descriptions. Guided by the extracted metadata, we synthesized a total of 15,207 model search requests with carefully designed prompts, as no such search requests are readily available. Experiments on the curated PTM dataset and the synthesized model search requests show that PTMPicker can help users effectively identify models,with 85% of the sampled requests successfully locating appropriate PTMs within the top-10 ranked candidates.
Recent advancements in artificial intelligence (AI) and its widespread integration into mobile software applications have received significant attention, highlighting the growing prominence of AI capabilities in modern software systems. However, the inherent hallucination and reliability issues of AI continue to raise persistent concerns. Consequently, application users and regulators increasingly ask critical questions such as: Does the application incorporate AI capabilities? and What specific types of AI functionalities are embedded? Preliminary efforts have been made to identify AI capabilities in mobile software; however, existing approaches mainly rely on manual inspection and rule-based heuristics. These methods are not only costly and time-consuming but also struggle to adapt advanced AI techniques. To address the limitations of existing methods, we propose LLMAID (Large Language Model for AI Discovery). LLMAID includes four main tasks: (1) candidate extraction, (2) knowledge base interaction, (3) AI capability analysis and detection, and (4) AI service summarization. We apply LLMAID to a dataset of 4,201 Android applications and demonstrate that it identifies 242% more real-world AI apps than state-of-the-art rule-based approaches. Our experiments show that LLM4AID achieves high precision and recall, both exceeding 90%, in detecting AI-related components. Additionally, a user study indicates that developers find the AI service summaries generated by LLMAID to be more informative and preferable to the original app descriptions. Finally, we leverage LLMAID to perform an empirical analysis of AI capabilities across Android apps. The results reveal a strong concentration of AI functionality in computer vision (54.80%), with object detection emerging as the most common task (25.19%).
Large Language Models (LLMs) have enabled the emergence of LLM agents, systems capable of pursuing under-specified goals and adapting after deployment. Evaluating such agents is challenging because their behavior is open ended, probabilistic, and shaped by system-level interactions over time. Traditional evaluation methods, built around fixed benchmarks and static test suites, fail to capture emergent behaviors or support continuous adaptation across the lifecycle. To ground a more systematic approach, we conduct a multivocal literature review (MLR) synthesizing academic and industrial evaluation practices. The findings directly inform two empirically derived artifacts: a process model and a reference architecture that embed evaluation as a continuous, governing function rather than a terminal checkpoint. Together they constitute the evaluation-driven development and operations (EDDOps) approach, which unifies offline (development-time) and online (runtime) evaluation within a closed feedback loop. By making evaluation evidence drive both runtime adaptation and governed redevelopment, EDDOps supports safer, more traceable evolution of LLM agents aligned with changing objectives, user needs, and governance constraints.
Before deploying an AI system to replace an existing process, it must be compared with the incumbent to ensure improvement without added risk. Traditional evaluation relies on ground truth for both systems, but this is often unavailable due to delayed or unknowable outcomes, high costs, or incomplete data, especially for long-standing systems deemed safe by convention. The more practical solution is not to compute absolute risk but the difference between systems. We therefore propose a marginal risk assessment framework, that avoids dependence on ground truth or absolute risk. It emphasizes three kinds of relative evaluation methodology, including predictability, capability and interaction dominance. By shifting focus from absolute to relative evaluation, our approach equips software teams with actionable guidance: identifying where AI enhances outcomes, where it introduces new risks, and how to adopt such systems responsibly.
The emergence of foundation models (FMs) has enabled the development of highly capable and autonomous agents, unlocking new application opportunities across a wide range of domains. Evaluating the architecture of agents is particularly important as the architectural decisions significantly impact the quality attributes of agents given their unique characteristics, including compound architecture, autonomous and non-deterministic behaviour, and continuous evolution. However, these traditional methods fall short in addressing the evaluation needs of agent architecture due to the unique characteristics of these agents. Therefore, in this paper, we present AgentArcEval, a novel agent architecture evaluation method designed specially to address the complexities of FM-based agent architecture and its evaluation. Moreover, we present a catalogue of agent-specific general scenarios, which serves as a guide for generating concrete scenarios to design and evaluate the agent architecture. We demonstrate the usefulness of AgentArcEval and the catalogue through a case study on the architecture evaluation of a real-world tax copilot, named Luna.
Federated foundation models represent a new paradigm to jointly fine-tune pre-trained foundation models across clients. It is still a challenge to fine-tune foundation models for a small group of new users or specialized scenarios, which typically involve limited data compared to the large-scale data used in pre-training. In this context, the trade-off between personalization and federation becomes more sensitive. To tackle these, we proposed a bi-level personalization framework for federated fine-tuning on foundation models. Specifically, we conduct personalized fine-tuning on the client-level using its private data, and then conduct a personalized aggregation on the server-level using similar users measured by client-specific task vectors. Given the personalization information gained from client-level fine-tuning, the server-level personalized aggregation can gain group-wise personalization information while mitigating the disturbance of irrelevant or interest-conflict clients with non-IID data. The effectiveness of the proposed algorithm has been demonstrated by extensive experimental analysis in benchmark datasets.
Foundation models, such as large language models (LLMs), have the potential to streamline evaluation workflows and improve their performance. However, practical adoption faces challenges, such as customisability, accuracy, and scalability. In this paper, we present DOCUEVAL, an AI engineering tool for building customisable DOCUment EVALuation workflows. DOCUEVAL supports advanced document processing and customisable workflow design which allow users to define theory-grounded reviewer roles, specify evaluation criteria, experiment with different reasoning strategies and choose the assessment style. To ensure traceability, DOCUEVAL provides comprehensive logging of every run, along with source attribution and configuration management, allowing systematic comparison of results across alternative setups. By integrating these capabilities, DOCUEVAL directly addresses core software engineering challenges, including how to determine whether evaluators are "good enough" for deployment and how to empirically compare different evaluation strategies. We demonstrate the usefulness of DOCUEVAL through a real-world academic peer review case, showing how DOCUEVAL enables both the engineering of evaluators and scalable, reliable document evaluation.