TVR: Automotive System Requirements T raceability V alidation and R ecovery Through Retrieval-Augmented Generation

Feifei Niu, Rongqi Pan, Lionel C. Briand, Hanyang Hu · 2025

Read TVR: Automotive System Requirements T raceability V alidation and R ecovery Through Retrieval-Augmented Generation by Feifei Niu, Rongqi Pan, Lionel C. Briand, Hanyang Hu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-15427.txt.

Introduction

Feifei Niu, Rongqi Pan, Lionel Briand, and Hanyang Hu present a 2025 tool that checks and repairs the links between car software requirements. Each system requirement should trace back to a stakeholder need, a link that proves the build matches what was asked and meets the rules. Change and human error often break or drop those links, which raises cost and risk.

Their tool pairs a large language model with a retrieval step that pulls in the most relevant requirements as context. It then judges whether a stated link is sound and hunts for links that are missing. Past tools did not tackle this stakeholder-to-system level, nor were they tried on real industry data. Tests in an industrial setting show the tool works well in practice. The team frames it as a strong aid for complex car programs.

Catalog record

CREATORFeifei Niu, Rongqi Pan, Lionel C. Briand, Hanyang Hu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE85 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-15427

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TVR: Automotive System Requirements T raceability V alidation and R ecovery Through Retrieval-Augmented Generation Feifei Niu, Rongqi Pan, Lionel C. Briand, Hanyang Hu Abstract In automotive software development, as well as other domains, traceability between stakeholder requirements and system requirements is crucial to ensure consistency, correctness, and regulatory compliance. However, erroneous or missing traceability relationships often arise due to improper propagation of requirement changes or human errors in requirement mapping, leading to inconsistencies and increased maintenance costs. Existing approaches do not address traceability between stakeholder and system requirements and are not validated on industrial data, where engineers manually establish the links between requirements. Additionally, there are variations in how requirements are expressed, posing challenges for training-based approaches, particularly in large-scale and heterogeneous automotive systems. Recent advancements in large language models (LLMs) provide new opportunities to address these challenges. In this paper, we introduce TVR, a requirement T raceability V alidation and R ecovery approach primarily targeting automotive systems, leveraging LLMs enhanced with retrieval-augmented generation (RAG). TVR is designed to validate existing traceability links and recover missing ones with high accuracy. We empirically evaluate TVR on real automotive requirements, achieving 98.87% accuracy in traceability validation and 85.50% correctness in traceability recovery. Additionally, TVR demonstrates strong robustness, achieving 97.13% in accuracy when handling unseen requirement variations. The experimental results highlight the practical effectiveness of TVR in industrial settings, offering a promising solution for improving requirements traceability in complex automotive systems. keywords Requirements Engineering, Traceability, Large Language Models 1 Introduction Automotive systems have become increasingly complex, comprising a wide range of integrated components and features, including powertrain, chassis, infotainment, advanced driver assistance, and electric and autonomous driving systems. These components collaborate through electronic control units (ECUs), sensors, actuators, and software, working seamlessly to deliver performance, safety, and user experience Wang et al. (2024) . To ensure high-quality software development, the automotive industry adheres to the Automotive Software Process Improvement and Capability dEtermination (ASPICE) standard Idri and Cheikhi (2016) ; García-Mireles et al. (2012) , which is based on the ISO 3300x norm group ISO/IEC 33001 (2015) ; ISO/IEC 33002 (2015) ; ISO/IEC 33003 (2015) ; ISO/IEC 33004 (2015) . A fundamental aspect of ASPICE-compliant software development is requirements engineering, in which requirements are defined at multiple levels, including stakeholder and system requirements. System requirements are derived from stakeholder requirements, with traceability links established to ensure consistency and correctness, especially amid frequent changes. Maintaining accurate and high-quality traceability links between stakeholder and system requirements is indeed crucial for multiple reasons: 1) Improving system consistency : Consistent traceability ensures that system requirements accurately reflect stakeholder needs, thus minimizing functional errors and inconsistencies Wiegers and Beatty (2013) ; Pargaonkar (2023) . 2) Preventing error propagation : Detecting incorrect requirement mappings early reduces maintenance cost overheads and avoids large-scale rework Tufail et al. (2017) . 3) Ensuring regulatory compliance , thus complying with ASPICE and functional-safety standards, including ISO 26262 Siegl et al. (2010) , to help improve the safety, reliability, and auditability of automotive software Qusef et al. (2011) . However, system engineers have reported challenges regarding traceability between requirements. One specific, particularly critical type of requirement we focus on here concerns Diagnostic Trouble Codes (DTCs), which are standardized vehicle error codes indicating malfunctions detected by the diagnostic system Marscholik and Subke (2009) ; Pirasteh et al. (2019) . While automotive diagnostic tools—with the market valued at USD 38.45 billion in 2023 and projected to reach USD 56.07 billion by 2031—play a crucial role in early issue detection, maintaining correct traceability of DTC requirements remains largely manual and labor-intensive. This limits the overall efficiency of the diagnostic process due to the time and effort required to manually verify the consistency between requirements and trace the system faults to requirements. According to system engineers, these challenges, primarily due to resource and time constraints, include: 1) Improper propagation of requirement changes : When stakeholder requirements are modified (e.g., new requirements are added or existing ones are updated), the corresponding system requirements may not be updated accordingly, leading to outdated or invalid traceability links. If a requirement is deleted or merged, the original traceability link may persist, leading to incorrect mappings and potentially causing functional inconsistencies or safety risks, resulting in irrecoverable losses in safety-critical domains. For example, in 1998, NASA’s $327 M Mars Climate Orbiter 1 1 1 https://en.wikipedia.org/wiki/Mars_Climate_Orbiter failed because mission directives mandated SI units, but the supporting software output imperial units. Without end-to-end traceability, the mismatch went undetected, causing trajectory errors and mission loss. This underscores the need for automated, robust requirement-consistency validation in industrial systems. 2) Errors in requirement mapping : Due to human error, system requirements may be incorrectly mapped to unrelated stakeholder requirements, or traceability links may be missing. According to our industry partner, they spend several weeks each year manually validating the traceability and consistency of requirements, as ensuring compliance with stakeholder requirements is always a top priority. Automating this process could reduce their manual effort and associated labor costs. Unlike similarity-based approaches Gao et al. (2022) ; Hey et al. (2021) ; Panichella et al. (2013) ; Kuang et al. (2015) ; Antoniol et al. (2002) ; Marcus and Maletic (2003) ; Keim et al. (2024) ; Keim et al. (2023) ; Cleland-Huang et al. (2005) that reconstruct links between requirements from scratch, industrial settings must manage preexisting traceability links, which, due to labeling errors or version changes, can incorrectly tie together unrelated or diverged requirements. Hence, it is essential to validate these links to ensure the connected requirements remain semantically and content-wise consistent. The goal of this paper is to provide automated, effective ways to address the challenges above and thus identify invalid and missing traceability links between DTC requirements . Because requirements are not always expressed consistently within or across projects, we leverage large language models (LLMs), for their demonstrated robustness to linguistic variation Brown et al. (2020) ; Qin et al. (2023) . Although prompt‐based LLM approaches have been recently applied to traceability recovery Fuchß et al. (2025) ; Rodriguez et al. (2023) , they either rely solely on simple Zero-Shot or Chain-of-Thought (CoT) prompts (e.g., “Is there a traceability link?”) for validating traceability links, which are inadequate for precisely validating the links between DTC system and stakeholder requirements that tend to slightly differ in the character strings of messages and signals. To address the traceability challenges described above in the DTC requirements, we propose TVR, a Retrieval-Augmented Generation (RAG) approach that leverages generative LLMs to support traceability validation and recovery. Unlike existing RAG-based traceability recovery approaches (e.g., Fuchß et al. (2025) ; Rodriguez et al. (2023) ), which primarily rely on retrieval to compute similarity scores between software artifacts to determine whether a traceability link exists, TVR adopts a fundamentally different approach to RAG. Specifically, TVR retrieves similar requirement pairs, including both positive (valid) and negative (invalid) traceability examples, and incorporates them into the prompt to explicitly guide and “teach” the LLM how to reason about traceability correctness. While traditional RAG systems are primarily designed for knowledge-intensive tasks, where external information must be retrieved from documents, knowledge bases, or repositories to provide necessary domain knowledge in the language model Lewis et al. (2020) , requirements traceability validation

Read the complete text