EFGAI 2026 Abstracts

Area 1 - EFGAI
Full Papers Short Papers

Area 1 - EFGAI

Full Papers
Paper Nr: 6
Title:

From Expert Reason Transitions to Local Reason Layers: A Small-Scale Feasibility Study toward AI-Ready Skill Transfer and Agentic Governance
Authors:

Masao Hirofuji and Masatoshi Hirofuji
Abstract: AI systems increasingly enter professional work, yet useful judgments often depend on reasons developed within particular workplaces and practices. We examine whether such locally situated reasoning can be reused by AI through Before-Reason-After (BRA) transitions, which represent a prior state, a reason associated with change, and the resulting state. Using longitudinal records from a bounded, expert-led consulting environment, we conducted a retrospective feasibility study in which BRA-derived ref-erences from prior cases were used to audit five held-out proposal-review cases without access to later expert reviews. Of 13 primary reasons later identified by the expert, 10 had Direct or Related antecedents in prior cases. Of seven residual expert reasons repre-sented in the operational reference inventory, both the reference-guided and generic au-ditors produced Full or Partial matches for five. Ten of the 13 expert reasons appeared in participants' immediate reflections, and five also appeared in later actual proposals. These observations support the feasibility of reusing BRA-derived local reason refer-ences across cases, without establishing superiority over generic auditing. They moti-vate an evolving Local Reason Layer built from repeated operational transitions. Larger longitudinal and cross-environment studies are needed to assess general skill-transfer effectiveness and agentic-governance applications.
Paper Nr: 7
Title:

Towards AI-Assisted Engineering and Formal Verification of Safety-Critical PLC-Based Railway Signalling Systems
Authors:

Christophe Ponsard and Antoine Ponsard
Abstract: Generative AI (GenAI) offers new opportunities for engineering industrial control systems (ICS), but their use in safety-critical contexts raises concerns about correctness, domain knowledge, traceability and human oversight. This paper proposes a model-based approach that combines LLM-based agents for PLC design and code generation with formal verification for correctness assurance. The approach uses IEC 61131 control models, targeting Ladder Diagrams. GenAI agents support requirements analysis, control-model generation and iterative refinement, whereas an external formal verification tool provides an independent mechanism for checking specified system properties. Human engineers remain responsible for requirements validation, review and approval of generated artefacts. The approach is investigated through a railway signalling system implemented on a realistic physical 1/8-scale demonstrator composed of about 40 blocks across 4 intertwined loops. We report on early experiments with contemporary LLM and agentic technologies starting from requirements to generate a PLC model and code, combined with formal verification guardrails. Based on these initial results, we also discuss more general topics such as appropriate levels of LLM specialisation, human-in-the-loop requirements and challenges for scaling the methodology.
Short Papers
Paper Nr: 8
Title:

Multi-Agent System for Automated Extraction and Structuring of Cybersecurity Requirements from Regulatory Documents
Authors:

Parth Sharma and Nika Strem
Abstract: Industrial control systems (ICS) and other operational technology (OT) assets are increasingly targeted by sophisticated cyberattacks. Guidelines for protecting these environments are codified by stringent regulatory requirements (e.\,g., IEC~62443, NERC~CIP, NIST~800~82, or ISO~27019) and internal safety policies. Compliance with these regulations is mandatory for reliable operation, yet in practice, compliance is achieved through labor‑intensive audits in which engineers must read extensive regulatory documents, interpret ambiguous clauses and manually translate the resulting requirements into technical controls and audit artifacts. This manual workflow suffers from inconsistent interpretation, limited scalability, and slow and expensive adaptation to evolving standards, leading to inefficiencies and heightened risk of regulatory penalties and security gaps. Existing approaches to the problem either rely on brittle rule‑based extraction or treat the problem as a monolithic task with limited generalization and oversight. To fill this gap, we propose a modular, agent‑based architecture that automates the end‑to‑end transformation of unstructured regulatory texts into machine‑interpretable specifications and executable compliance artifacts. A coordinated suite of specialized agents performs document digitization, requirement extraction, semantic structuring, and generation of actionable compliance code, while an orchestrator enables human‑in‑the‑loop integration, enabling scalability and preventing error propagation. An interactive web interface supports configurable extraction strategies and review of extracted requirements, preserving expert oversight while reducing manual effort. The envisioned system enables the transition of compliance assurance from a manual activity to a semi‑automated, auditable process, aiming for fast, consistent and scalable compliance with cybersecurity regulations in industrial environments.