heuristic development

This study addresses the challenge of isolated analysis of multi-source heterogeneous data—such as work order texts, sensor signals, and fault knowledge—in industrial maintenance, which often hinders explainable condition-based decision-making. To overcome this limitation, the authors propose the Condition Insight Agent framework, which integrates maintenance language, abstracted operational behaviors, and structured fault semantics. By leveraging deterministic evidence construction and a rule-driven validation loop, the framework enables constrained yet reliable reasoning with large language models (LLMs). Designed to operate under human oversight, the approach effectively handles data sparsity and heterogeneity while generating interpretable, evidence-backed maintenance recommendations. Its reliability and practicality have been validated through real-world deployment in a Computerized Maintenance Management System (CMMS).

This study addresses the prevalent yet elusive issue of “silent failures” in AI-generated code—functional degradations or errors that appear syntactically correct but evade detection by conventional testing methods. The authors propose the Reward Shaping Failure Hypothesis, attributing this phenomenon to optimization dynamics in human feedback alignment, and introduce the novel concept of “failure truthfulness.” Building on this insight, they develop AIRA, the first targeted risk auditing framework for AI-generated code, which employs structured static analysis with 15 deterministic rules across JavaScript, Python, and TypeScript, with particular emphasis on exception handling. Empirical evaluation reveals that AI-generated code exhibits 1.80 times more high-severity issues per file (0.435) than human-written code (0.242), a consistent effect across languages, with deficiencies in exception handling being especially pronounced.

This study addresses the challenge that defeaters in safety arguments—due to their unstructured descriptions and lack of standardized representation—are difficult to review, trace, and reuse. To resolve this, the work proposes Defeater Cards, a novel standardized documentation artifact grounded in the 5W1H framework, offering the first systematic formalism for representing defeaters. The card structure was developed through a literature review and thematic analysis, and its efficacy was validated across multiple case studies spanning diverse domains. Empirical results demonstrate that Defeater Cards effectively expose implicit assumptions and reasoning gaps, substantially enhancing the auditability, traceability, and evolvability of safety arguments. An open-source repository of Defeater Cards is also released to foster knowledge reuse and community-driven collaboration.

Existing AI-generated content (AIGC) fact-checking tools predominantly rely on black-box binary classification or regression models, suffering from poor interpretability, limited evidence diversity, and minimal user interactivity. Method: We propose the first user-driven, fine-grained fact verification framework that decomposes long texts into atomic claims, integrates heterogeneous multi-source evidence (e.g., knowledge bases, web pages, documents), and employs cross-source evidence fusion with an interpretable reasoning model to produce claim-level confidence scores and natural-language explanations—supporting multi-hop provenance tracing and dynamic user feedback. Contribution/Results: Our framework breaks from conventional paradigms by enabling transparent, traceable, evidence-diverse, and human-AI collaborative verification. Experiments demonstrate significant improvements in user verification efficiency (+37%) and trust (+42%), establishing a novel paradigm for trustworthy AIGC interaction.

This study addresses the challenge of deep neural networks failing to proactively identify erroneous predictions—such as false negatives—in industrial defect detection, which compromises reliability in safety-critical applications. The authors propose a novel post-hoc explanation–based reliability metric that quantifies the difference in Intersection over Union (IoU) between class-specific and class-agnostic discriminative saliency maps as a reliability score. To amplify this discrepancy and enhance error detection, adversarial augmentation is introduced. This approach enables the first active identification of potentially erroneous model outputs in binary classification tasks, establishing a new “data–model–explanation–output” paradigm that significantly improves the trustworthiness of deployed AI systems. Evaluated on two industrial defect detection benchmarks, the method achieves 100% recall for false negatives when combined with adversarial augmentation.

This study addresses a critical gap in understanding whether large language models (LLMs) are susceptible to human-like cognitive heuristics—such as the halo effect, framing effect, and anchoring bias—in the context of code vulnerability detection. The authors introduce a controlled experimental framework that manipulates contextual cues while keeping the underlying code fixed, enabling cross-model and cross-language evaluation across eight prominent LLMs and three programming languages. Combining quantitative analysis with fine-grained, code-level inspection, they find that all models exhibit significant susceptibility to cognitive biases, with an average 33.2% performance drop due to framing effects. Semantic vulnerabilities prove especially prone to heuristic interference, and models frequently misjudge code safety without accurately localizing flaws. Leveraging these insights, the authors propose a cognitive attack method capable of suppressing 97% of detected vulnerabilities, exposing fundamental weaknesses and suggesting avenues for robustness improvement.

Current safety and assurance arguments for AI-driven systems often lack genuine intelligibility, hindering responsible engineering and governance decisions. Building on Catherine Elgin’s epistemology of understanding, this work operationalizes “understanding” within the Assurance 2.0 framework by introducing two interrelated artifacts: “understanding bases” and “individual understanding statements.” Through structured argumentation, evidence modeling, confidence assessment, and defeater mechanisms, the approach explicitly articulates system behavior, underlying assumptions, and uncertainties, while highlighting how automation may inadvertently erode human understanding. By providing developers, assessors, and decision-makers with verifiable and contestable grounds for understanding, this method establishes an initial pathway for evaluating both the efficacy of understanding and its cognitive impacts, thereby advancing practices of responsible AI governance.

This work addresses the overwhelming volume of alerts generated by static analysis tools, which often exceeds human review capacity. To tackle this challenge, the authors propose an automated triage approach leveraging large language models (LLMs), integrating consistency checking (CC) with LLM-based reasoning evaluation (LRE) to substantially reduce false positives. The method further incorporates dynamic triggering to validate the exploitability of reported vulnerabilities. Evaluated on medium-scale LLMs—including o4-mini and gpt-oss-120b—across the Juliet, FormAI, and SV-COMP benchmarks, the approach achieves a recall of at least 98% and a specificity of no less than 94.8%, demonstrating significantly improved triage accuracy and generalization capability.

This work addresses the prevalent yet often undetectable issue of logical inconsistency between reasoning and final answers in chain-of-thought (CoT) outputs generated by current AI systems during safety evaluations. The study is the first to formally distinguish between reasoning consistency and faithfulness, introducing a taxonomy encompassing six distinct types of inconsistency. To enable post-hoc detection without modifying model generation, the authors propose InspectScout—a reusable scanning method grounded in formal definitions, supported by a human-annotated benchmark, and implemented via an automated detection algorithm integrated into the inspect_evals framework. Experiments demonstrate that reasoning inconsistencies are widespread across four mainstream models and three safety-related tasks, and can be reliably identified; moreover, the patterns of such inconsistencies exhibit systematic variation across models.

Current AI regulation lacks a quantifiable definition of “acceptable risk” and corresponding verification mechanisms, hindering compliance certification for high-risk systems. This work proposes the first two-stage statistical certification framework for black-box AI: regulatory authorities first define an acceptable failure probability and input domain; then, leveraging RoMA/gRoMA methods, the framework computes an auditable upper bound on the failure rate for AI systems of arbitrary architecture without requiring access to internal model details. By adapting aviation safety certification paradigms to AI governance, the approach employs statistical hypothesis testing and scalable black-box verification to generate legally defensible safety certificates. It directly aligns with regulatory requirements such as the EU AI Act and shifts accountability upstream to developers.