Document Retrieval Augmented Fine-Tuning (DRAFT) for safety-critical software assessments

Safety critical software assessment requires robust assessment against complex regulatory frameworks, a process traditionally limited by manual evaluation. This paper presents Document Retrieval-Augmented Fine-Tuning (DRAFT), a novel approach that enhances the capabilities of a large language model (LLM) for safety-critical compliance assessment. DRAFT builds upon existing Retrieval-Augmented Generation (RAG) techniques by introducing a novel fine-tuning framework that accommodates our dual-retrieval architecture, which simultaneously accesses both software documentation and applicable reference standards. To fine-tune DRAFT, we develop a semi-automated dataset generation methodology that incorporates variable numbers of relevant documents with meaningful distractors, closely mirroring real-world assessment scenarios. Experiments with GPT-4o-mini demonstrate a 7% improvement in correctness over the baseline model, with qualitative improvements in evidence handling, response structure, and domain-specific reasoning. DRAFT represents a practical approach to improving compliance assessment systems while maintaining the transparency and evidence-based reasoning essential in regulatory domains.

Multi-Stage Retrieval for Operational Technology Cybersecurity Compliance Using Large Language Models: A Railway Casestudy

Operational Technology Cybersecurity (OTCS) continues to be a dominant challenge for critical infrastructure such as railways. As these systems become increasingly vulnerable to malicious attacks due to digitalization, effective documentation and compliance processes are essential to protect these safety-critical systems. This paper proposes a novel system that leverages Large Language Models (LLMs) and multi-stage retrieval to enhance the compliance verification process against standards like IEC 62443 and the rail-specific IEC 63452. We first evaluate a Baseline Compliance Architecture (BCA) for answering OTCS compliance queries, then develop an extended approach called Parallel Compliance Architecture (PCA) that incorporates additional context from regulatory standards. Through empirical evaluation comparing OpenAI-gpt-4o and Claude-3.5-haiku models in these architectures, we demonstrate that the PCA significantly improves both correctness and reasoning quality in compliance verification. Our research establishes metrics for response correctness, logical reasoning, and hallucination detection, highlighting the strengths and limitations of using LLMs for compliance verification in railway cybersecurity. The results suggest that retrieval-augmented approaches can significantly improve the efficiency and accuracy of compliance assessments, particularly valuable in an industry facing a shortage of cybersecurity expertise.