Pith. sign in

REVIEW 4 major objections 6 minor 76 references

Survey of GenAI for Automotive Software Development: From Requirements to Executable Code

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Commercial large language models are entering automotive software development through code generation, while requirements handling lags because OEM requirements are confidential and must stay off third-party cloud services.

desk verdict Readable survey that maps the field and names a plausible bottleneck, but its quantitative trends rest on a non-systematic, self-skewed corpus and a 9-person survey. read the letter →

arxiv 2507.15025 v1 pith:O3LHV3YU submitted 2025-07-20 cs.SE cs.AI

Nenad Petrovic , Vahid Zolfaghari , Andre Schamschurko , Sven Kirchner , Fengjunjie Pan , Chengdng Wu , Nils Purschke , Aleksei Velsh
show 5 more authors
Krzysztof Lebioda Yinglei Song Yi Zhang Lukasz Mazur Alois Knoll
This is my paper
classification cs.SEcs.AI
keywords automotivesoftwaredevelopmentgenerativeAIlargelanguagemodelsretrieval-augmentedgenerationvision-languagecoderequirementsengineeringcompliance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This survey of about sixty published works plus a nine-respondent industry questionnaire tries to pin down where generative AI has actually entered automotive software development. It finds that code generation is the clearest adoption scenario, with commercial models such as GPT-4 still dominant, and that requirements handling is the lagging step because requirements are protected by non-disclosure agreements and companies avoid third-party cloud services. The paper also derives a generalized GenAI-aided workflow that runs from RAG-indexed documents, through formal models, to simulation and target code, and reviews the prompting, retrieval, vision-language, and hallucination-mitigation techniques used along the way. If its sample is representative, the paper provides a single map of where GenAI can be adopted now and where the bottleneck to wider adoption sits.

What carries the argument

The central object is the generalized GenAI-aided automotive software development workflow, a pipeline that chains RAG-based chunking and indexing of requirements and regulation documents, VLM extraction of visual information, LLM summarization into formal model representations such as Ecore metamodels, XMI instances, and JSON schemas, design-time compliance checks, LLM code generation for simulation and target platforms, and a human reviewer with feedback from static analysis and runtime profiling. The mechanism that explains the adoption pattern is the confidentiality split: open, specification-driven code tasks are served well by commercial GPT-family models, while document-bound requirements tasks demand locally deployable fine-tuned models, whose adoption is constrained by the lack of public automotive datasets. That split, the paper argues, is why requirements handling is the main bottleneck rather than the models' raw capability.

What would settle it

A reader could independently rerun the literature search under explicit inclusion and exclusion criteria with fixed databases and date ranges, recompute the model-usage histograms, and survey a larger set of automotive companies; if the distribution stopped showing GPT-family dominance in code generation, or showed substantial use of third-party GenAI services on requirements, the paper's central adoption-and-bottleneck claim would be contradicted.

Watch

Extended reading notes

Core claim

The paper's central claim is descriptive: as of 2024-2025 the automotive industry is at the beginning of GenAI adoption, and adoption is uneven along the development chain. In code-generation tasks, such as simulator code, vehicle configuration scripts, and safety-relevant C code, commercial GPT-family models are the dominant tools in the reviewed literature. Requirements handling and compliance are less automated: the preferred pattern there is locally deployable, fine-tuned models such as Llama 3, because OEM requirements are confidential and cannot be exposed to third-party services, but the shortage of public automotive datasets makes fine-tuning hard. From the reviewed works the paper derives a generalized GenAI-aided workflow that chains RAG-indexed documents, VLM extraction of diagrams, LLM formalization into models such as Ecore and JSON, design-time compliance checks, LLM code generation, simulation, and human review, and it presents a nine-person industry survey in which nearly everyone uses GenAI for code, only one person for requirements, and most tools are commercial services.

Load-bearing premise

The load-bearing premise is that the roughly sixty curated papers and nine industry survey responses fairly represent automotive industry practice, since the search is human-filtered without stated inclusion criteria and a noticeable share of the reviewed entries are the authors' own project works; if the sample is unrepresentative, the adoption trends and the generalized workflow would lose their evidence base.

Editorial extensions

If this is right

  • Automotive code generation will keep depending on commercial GPT-family models until local models match their output quality on open specification-driven tasks.
  • Requirements and user-story handling will remain the least automated stage unless locally deployable fine-tuned models become practical, because confidentiality rules out third-party cloud processing for most OEMs.
  • The generalized workflow gives tool vendors and OEMs a concrete target: invest in RAG indexing, VLM diagram extraction, and formal model integration to move GenAI adoption beyond code generation.
  • Prompting practice will likely split by task, with chain-of-thought prompting for behavior-related code and specification-driven prompting for simulation scenarios, test cases, and vehicle configuration scripts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the confidentiality split points to synthetic or de-identified automotive requirements datasets, or federated fine-tuning, as the most direct route to closing the requirements-handling gap, since real requirements cannot leave corporate boundaries.
  • A testable extension would be to benchmark generated safety-relevant code against MISRA and ISO 26262 checks and count how often a human reviewer must intervene, which would convert the survey's qualitative adoption story into a cost-benefit model.
  • The workflow's division of labor, with commercial models for code and local models for documents, could be adopted immediately by an engineering organization even before local fine-tuning quality catches up, because both sides of the split already exist in practice.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This manuscript surveys the use of generative AI—large language models (LLMs), retrieval-augmented generation (RAG), vision-language models (VLMs), and prompting techniques—across automotive software development, with emphasis on requirements handling, regulation compliance, and code generation. It presents several tables of relevant works, histograms of model and prompting-strategy usage, a generalized workflow from requirements to executable code, a discussion of hallucination-mitigation methods, and results of a small survey of industry partners. The main conclusions are that the automotive industry is beginning to adopt GenAI tools, that commercial GPT-family models dominate code generation, and that requirements handling remains an adoption gap because of NDA-related constraints that favor locally deployable fine-tuned models.

Significance. If accepted, this would be one of the few integrative overviews of an emerging area, providing a broad map of GenAI techniques, tools, and challenges in automotive software development. The paper's strengths are its topical breadth (LLMs, RAG, VLMs, prompting strategies, hallucination mitigation), its explicit acknowledgement of industrial constraints such as NDAs, and its carefully hedged wording of the conclusions. However, the evidentiary basis for the quantitative trends is currently weak: the corpus is curated without documented selection criteria, the tables and histograms were produced with LLM-based extraction without described verification, and a substantial share of the evidence rows come from the authors' own project. The paper would be a useful resource after these data-provenance and self-citation issues are addressed.

major comments (4)
  1. [Section V / Table I / Section XIII] The paper's load-bearing claim about requirements handling is not supported by its own Table I. Section V states that 'for this type of tasks, locally deployable, fine-tuned models are preferable (such as fine-tuned Llama 3)', and Section XIII concludes that 'for tasks directly handling requirements, fine-tuned and optimized variants of locally deployable models are used, especially Llama3'. Yet six of the seven rows in Table I include commercial GPT-family models (GPT-4o, GPT-4 Turbo, GPT-4, o1-preview), and Llama appears in only four rows. The table therefore shows a mix of commercial and local models, not a clear dominance of locally deployed fine-tuned Llama. This internal inconsistency affects the central conclusion in Section XIV and should be corrected or re-analyzed per row.
  2. [Section III] The methodology does not specify inclusion/exclusion criteria, a date range, a database search protocol, or a procedure for handling duplicates and quality, beyond stating that papers 'are curated by human reader and filtered'. Because the tables and histograms in Figs. 4 and 6 were generated by prompting Google NotebookLM with no described human verification of the extracted entries, the quantitative claims about model usage and prompting-strategy frequency cannot be independently checked. The authors should document the full search protocol, the filtering criteria, and a verification step for the extracted data.
  3. [Tables I, III, V, VI, VII] A substantial share of the in-scope entries are from the authors' own CeCaS-related works (e.g., references [16], [19], [21]-[25], [34]-[36], [55], [56], [59]); in Table I, six of seven rows are author-group works. Because the CeCaS program is explicitly built around locally deployable fine-tuned models and formal-model-based code generation, the evidence base is not independent of the conclusions. The paper should report this overlap, re-run the trend analyses (Figs. 4 and 6) with self-citations excluded, and state whether the qualitative conclusions in Section XIV survive that exclusion.
  4. [Section XII] The industry survey is based on only nine respondents, seven from industry, and no per-question response counts are reported; the text only gives qualitative summaries ('all of the participants', 'major part'). The paper subsequently uses this survey to support the conclusion that there is a gap in requirements handling. Please provide the full response distributions for each question, explicitly state that the respondents are participants from the CeCaS project, and label the survey as an illustrative convenience sample rather than evidence of industry-wide practice.
minor comments (6)
  1. [Section VIII.B] The heading 'Visual hallucinations' is followed by no content; either provide the subsection or remove the heading.
  2. [Section IV] The text says 'Fig. 1 depicts the generalized workflow', but Fig. 1 shows the methodology; the generalized workflow appears in Fig. 2.
  3. [Section VI] The sentence '1 is a synthesized general pipeline...' appears to contain a placeholder figure reference and should be corrected to an actual figure or removed.
  4. [Section XI] Reference [64] is duplicated in the sentence describing CodeT5; the citation should appear once.
  5. [Table VII] Table VII is titled 'Summary of Prompting Strategies in Automotive Use Cases' but includes several non-automotive works (e.g., [64], [69], [70]-[72]); the caption should clarify the table's scope or the non-automotive rows should be moved elsewhere.
  6. [Section XI] There is a typo in the phrase 'techqniues' that should read 'techniques'.

Circularity Check

3 steps flagged · score 4.0 of 10

Survey conclusions are partly self-referential: the requirements-handling 'preference' for locally deployable fine-tuned models and the GPT/Llama-dominance histograms are drawn from a corpus in which the authors' own CeCaS papers are heavily represented, and the generalized workflow is explicitly derived from that same project.

  1. self citation load bearing [Section V 'LLMs and RAG for Requirements Handling' and Table I]
    "Based on the existing works, it can be noticed that for this type of tasks, locally deployable, fine-tuned models are preferable (such as fine-tuned Llama 3), which was expected considering the non-disclosure policies of the underlying organizations. Table I summarizes relevant works, use cases and models."

    The 'existing works' supporting this claim are Table I rows, which are predominantly the authors' own CeCaS-related papers ([19], [21], [22], [23], [24], [25]). Those works were themselves built on the premise that OEM requirements are NDA-protected and therefore require local, fine-tuned models. The paper then presents that premise as an observed preference derived from the cited literature. The conclusion is thus contained in the self-citation cluster rather than established by independent external evidence.

  2. other [Section XIII 'LLM Usage Overview' and Fig. 6]
    "In Fig. 6, a summary of LLM usage for dominant models across various automotive-related use cases is given. According to the results, it can be noticed that solutions based on OpenAI's commercial GPT model are still dominant, especially the latest variants. ... for tasks directly handling requirements, fine-tuned and optimized variants of locally deployable models are used, especially Llama3 which is quite dominant among them."

    The histograms are counts over a corpus whose in-scope tables include many of the authors' own works (for example [22], [23], [24], [34], [36], [55], [56], [59]) that use GPT-4o/GPT-4 and Llama 3 as a project choice. Section III states the corpus is 'curated by human reader and filtered' and that tables/histograms are extracted via NotebookLM without described verification. The reported dominance of GPT and Llama is therefore, by construction, a property of the authors' own selection and model choices, not an independent industry-wide observation.

1 more flagged steps
  1. renaming known result [Section IV 'Generalized Workflow' and Fig. 2]
    "In this section, we present the generalized workflow of GenAI-empowered approach to automotive software development, based on both the existing literature and our 2-year work on CeCaS project [14] [15] [16]."

    The 'generalized workflow' is explicitly based on the CeCaS project, and it is then used to organize the reviewed literature and to motivate the conclusion that there is a research gap in requirements handling. The workflow steps (RAG chunking of regulations, VLM extraction, formal-metamodel intermediate representations, LLM code generation, human-in-the-loop) restate the CeCaS pipeline, so treating those steps as findings from the literature review partly renames the authors' own project architecture as a field-wide generalization.

full rationale

This is a survey rather than a formal derivation, so there is no equation-level circularity in the usual sense. The paper does contain independent content: external works are cited, an n=9 industry survey is reported, and the histograms at least describe whatever corpus was assembled. However, the central adoption narrative and the requirements-handling gap rest partly on a self-citation chain. Table I's requirements-handling rows are dominated by the authors' own CeCaS papers, and those same papers contribute to the GPT/Llama counts in Figs. 4 and 6. The claim that locally deployable fine-tuned models are 'preferable' for requirements handling is asserted immediately before citing those self-authored works, and the generalized workflow is explicitly derived from the same project. Because the corpus selection described in Section III is non-systematic and human-filtered, the self-skew is not corrected by an independent benchmark. This makes the survey's trend findings partially self-referential, but not fully circular, since the paper's claims still have content outside the self-citation cluster. Score 4 reflects partial circularity through self-citation and corpus construction.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The survey contributes no fitted parameters and no invented entities. Its findings rest on three unproved domain assumptions: representativeness of the hand-curated paper sample, reliability of NotebookLM-based table extraction, and representativeness of the 9-person survey. The first is the most load-bearing given the self-citation weight in the sample.

assumptions (3)
  • domain assumption The human-curated, non-systematic selection of roughly 60 papers represents the state of the art in GenAI for automotive software development.
    Section III states papers 'are curated by human reader and filtered' with no inclusion or exclusion criteria; the trend histograms in Figs. 4 and 6 inherit this selection, which includes many of the authors' own works.
  • domain assumption Google NotebookLM summarization reliably produces accurate tables and usage statistics from the collected PDFs.
    Section III describes uploading PDFs to NotebookLM and prompting it to extract tables and histograms; no human verification step is described.
  • domain assumption The 9-response partner questionnaire indicates automotive industry adoption patterns.
    Section XII bases conclusions such as 'all of the participants do rely on GenAI-based tools' on 9 respondents (1 university, 1 research institute, 7 industry); Section XIV generalizes to 'automotive industry has positive perception'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Survey of GenAI for Automotive Software Development: From Requirements to Executable Code." pith.science (2026). https://pith.science/paper/O3LHV3YU

@misc{pith2026250715025,
  author       = {Pith},
  title        = {Pith review of: Survey of GenAI for Automotive Software Development: From Requirements to Executable Code},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O3LHV3YU}},
  note         = {Machine review of arXiv:2507.15025}
}
read the original abstract

Adoption of state-of-art Generative Artificial Intelligence (GenAI) aims to revolutionize many industrial areas by reducing the amount of human intervention needed and effort for handling complex underlying processes. Automotive software development is considered to be a significant area for GenAI adoption, taking into account lengthy and expensive procedures, resulting from the amount of requirements and strict standardization. In this paper, we explore the adoption of GenAI for various steps of automotive software development, mainly focusing on requirements handling, compliance aspects and code generation. Three GenAI-related technologies are covered within the state-of-art: Large Language Models (LLMs), Retrieval Augmented Generation (RAG), Vision Language Models (VLMs), as well as overview of adopted prompting techniques in case of code generation. Additionally, we also derive a generalized GenAI-aided automotive software development workflow based on our findings from this literature review. Finally, we include a summary of a survey outcome, which was conducted among our automotive industry partners regarding the type of GenAI tools used for their daily work activities.

Figures

Figures reproduced from arXiv: 2507.15025 by the authors.

Figure 1
Figure 1. Underlying methodology for each of the topics covered used for this survey paper creation [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. GenAI-driven automotive software development workflow [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Summary of Scenario Generation solutions [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Histogram of Prompting Strategy Usage Across Re [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Usage of GenAI in automotive software development [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Histogram of LLM Usage Across Reviewed Works [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 50 canonical work pages

  1. [16]

    Synergy of large language model and model driven engineering for automated development of centralized vehicular systems,

    N. Petrovic et al. , “Synergy of large language model and model driven engineering for automated development of centralized vehicular systems,” Technical University of Munich, Technical Report, 2024. [Online]. Available: https://mediatum.ub.tum.de/doc/1738462/1738462. pdf

    arXiv 2024

  2. [19]

    Adopting rag for llm-aided future vehicle design,

    V . Zolfaghari, N. Petrovic, F. Pan, K. Lebioda, and A. Knoll, “Adopting rag for llm-aided future vehicle design,” in 2024 2nd International Conference on Foundation and Large Language Models (FLLM) , 2024, pp. 437–442

    work page 2024

  3. [21]

    Querying large automotive software models: Agentic vs. direct llm approaches,

    Łukasz Mazur, N. Petrovic, J. P. Miranda, A. Radermacher, R. Rasche, and A. Knoll, “Querying large automotive software models: Agentic vs. direct llm approaches,” pp. 1–8, June 2025, arXiv preprint arXiv:2506.13171. [Online]. Available: https://arxiv.org/abs/2506.13171

    arXiv 2025

  4. [25]

    Optimizing retrieval augmented generation for object constraint language,

    K. C. Li, V . Zolfaghari, N. Petrovic, F. Pan, and A. Knoll, “Optimizing retrieval augmented generation for object constraint language,” in arXiv preprint arXiv:2505.13129, May 2025, available at https://arxiv.org/abs/ 2505.13129

    arXiv 2025

  5. [34]

    Automating automotive software development: A synergy of generative ai and formal methods,

    F. Pan, Y . Song, L. Wen, N. Petrovic, K. Lebioda, and A. Knoll, “Automating automotive software development: A synergy of generative ai and formal methods,” pp. 1–8, May 2025, arXiv preprint arXiv:2505.02500. [Online]. Available: https://arxiv.org/abs/2505.02500

    arXiv 2025

  6. [36]

    Llm-driven testing for autonomous driving scenarios,

    N. Petrovic, K. Lebioda, V . Zolfaghari, A. Schamschurko, S. Kirchner, N. Purschke, F. Pan, and A. Knoll, “Llm-driven testing for autonomous driving scenarios,” in 2024 2nd International Conference on Foundation and Large Language Models (FLLM) , Dubai, United Arab Emirates, 2024, pp. 173–178

    work page 2024

  7. [55]

    Available: https://arxiv.org/abs/2308.07702

    [Online]. Available: https://arxiv.org/abs/2308.07702

    arXiv

  8. [56]

    Generating Automotive Code: Large Language Models for Software Development and Verification in Safety-Critical Systems

    S. Kirchner and A. Knoll, “Generating automotive code: Large language models for software development and verification in safety-critical systems,” in arXiv preprint arXiv:2506.04038 , June 2025, available at https://arxiv.org/abs/2506.04038

    work page Pith review arXiv 2025

  9. [59]

    Speedgen: Enhancing code ef- ficiency through large language model-based performance optimization,

    N. Purschke, S. Kirchner, and A. Knoll, “Speedgen: Enhancing code ef- ficiency through large language model-based performance optimization,” in 2025 IEEE International Conference on Software Analysis, Evolution and Reengineering (SANER) , Montreal, QC, Canada, 2025, pp. 1–12, available at https://doi.org/10.1109/SANER64311.2025.00045

    arXiv 2025

Show all 76 references
  1. [1]

    M. L. Baptista, N. Yue, M. M. Manjurul Islam, and H. Prendinger, Large Language Models (LLMs) for Smart Manufacturing and Industry X.0. Cham: Springer Nature Switzerland, 2025, pp. 97–119. [Online]. Available: https://doi.org/10.1007/978-3-031-80154-9 5

    2025 doi

  2. [2]

    Exploring generative ai in automated software engineering,

    M. Staron and S. Abrah ˜ao, “Exploring generative ai in automated software engineering,” IEEE Software, vol. 42, no. 3, pp. 142–145, 2025

    2025

  3. [3]

    Generative ai adoption in automotive vehicle technology: Case study of custom gpt,

    A. Phatale and A. Kaushik, “Generative ai adoption in automotive vehicle technology: Case study of custom gpt,” Journal of Artificial Intelligence & Cloud Computing , vol. 3, pp. 1–5, 11 2024

    2024

  4. [4]

    Auto- motive software engineering in an increasingly data-driven automotive sector,

    O. Denninger, J. K. Axmann, S. Kacianka, and B. Westphal, “Auto- motive software engineering in an increasingly data-driven automotive sector,” 2025

    2025

  5. [5]

    Requirements and software engineering for automotive perception systems: an interview study,

    K. M. Habibullah, H.-M. Heyn, G. Gay, J. Horkoff, E. Knauss, M. Borg, A. Knauss, H. Sivencrona, and P. J. Li, “Requirements and software engineering for automotive perception systems: an interview study,” Requirements Engineering , vol. 29, no. 1, pp. 25–48, Mar

  6. [6]

    Requirements management in automotive: Tools and trends for a competitive edge,

    M. Maier, “Requirements management in automotive: Tools and trends for a competitive edge,” January 2023, accessed: 2025-04-11. [Online]. Available: https://www.semantha.de/ requirements-management-in-automotive-tools-and-trends-for-a-competitive-edge/

    2023

  7. [7]

    A survey on large language models with some insights on their capabilities and limitations,

    A. Matarazzo and R. Torlone, “A survey on large language models with some insights on their capabilities and limitations,” 2025. [Online]. Available: https://arxiv.org/abs/2501.04040v1

    2025 arXiv

  8. [8]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert- V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Ches...

    2020 arXiv

  9. [9]

    Pre-train prompt fine-tune: A survey of prompting methods in natu- ral language processing,

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, N. Neubig, and G. Neubig, “Pre-train prompt fine-tune: A survey of prompting methods in natu- ral language processing,” 2021, available at https://arxiv.org/abs/2107. 13586

    2021

  10. [10]

    Using the Retrieval- Augmented Generation to Improve the Question-Answering System in Human Health Risk Assessment: The Development and Application,

    W. Meng, Y . Li, L. Chen, and Z. Dong, “Using the Retrieval- Augmented Generation to Improve the Question-Answering System in Human Health Risk Assessment: The Development and Application,” Electronics, vol. 14, no. 2, p. 386, Jan. 2025, number: 2 Publisher: Multidisciplinary ...

    2025

  11. [11]

    Automating regulatory compliance: A multi-agent solution using Amazon Bedrock and CrewAI | Artificial Intelligence and Machine Learning,

    “Automating regulatory compliance: A multi-agent solution using Amazon Bedrock and CrewAI | Artificial Intelligence and Machine Learning,” Apr. 2025, section: Amazon Bedrock. [Online]. Available: https://aws.amazon.com/blogs/machine-learning/ automating-regulatory-compliance-a...

    2025

  12. [12]

    Vision language models in autonomous driving: A survey and outlook,

    X. Zhou, M. Liu, E. Yurtsever, B. L. Zagar, W. Zimmer, H. Cao, and A. C. Knoll, “Vision language models in autonomous driving: A survey and outlook,” IEEE Transactions on Intelligent Vehicles, pp. 1–20, 2024

    2024

  13. [13]

    NotebookLM: Ai-powered research and note-taking assistant,

    Google, “NotebookLM: Ai-powered research and note-taking assistant,” https://notebooklm.google/, 2025, accessed: 2025-05-03

    2025

  14. [14]

    MANNHEIM-CeCaS – Central Car Server – Supercomputing for Automotive,

    Chair of Robotics, Artificial Intelligence and Real-time Systems, “MANNHEIM-CeCaS – Central Car Server – Supercomputing for Automotive,” n.d., accessed: 2025-05-02. [Online]. Available: https://www.ce.cit.tum.de/air/research/mannheim-cecas/

    2025

  15. [15]

    Towards single-system illusion in software-defined vehicles - automated, ai-powered workflow,

    K. Lebioda, V . V orobev, N. Petrovic, F. Pan, V . Zolfaghari, and A. Knoll, “Towards single-system illusion in software-defined vehicles - automated, ai-powered workflow,” Technical University of Munich, Technical Report, 2024. [Online]. Available: https: //mediatum.ub.tum.de...

    2024

  16. [17]

    Generative artificial intelligence for model-based graphical programming in automotive function development,

    A. Abdalla, H. Pandey, B. Shomali, J. Schaub, A. M ¨uller, M. Eisen- barth, and J. Andert, “Generative artificial intelligence for model-based graphical programming in automotive function development,” Novem- ber 10 2024, available at SSRN: https://ssrn.com/abstract=5153452 or...

    2024 doi

  17. [18]

    RECSIP: REpeated Clustering of Scores Improving the Precision,

    A. Schamschurko, N. Petrovic, and A. C. Knoll, “RECSIP: REpeated Clustering of Scores Improving the Precision,” 2025, conference paper accepted for IntelliSys2025. [Online]. Available: https://arxiv.org/abs/2503.12108

    2025 arXiv

  18. [20]

    Local large language models to simplify requirement engineering documents in the automotive industry,

    Y . Uygun and V . Momodu, “Local large language models to simplify requirement engineering documents in the automotive industry,” Production & Manufacturing Research , vol. 12, no. 1, 2024. [Online]. Available: https://doi.org/10.1080/21693277.2024.2375296

    2024

  19. [22]

    Llm-based iterative approach to metamodeling in automotive,

    N. Petrovic, F. Pan, V . Zolfaghari, and A. Knoll, “Llm-based iterative approach to metamodeling in automotive,” 2025

    2025

  20. [23]

    Llm-enabled instance model generation,

    F. Pan, N. Petrovic, V . Zolfaghari, L. Wen, and A. Knoll, “Llm-enabled instance model generation,” 2025

    2025

  21. [24]

    Generative ai for ocl constraint generation: Dataset collection and llm fine-tuning,

    F. Pan, V . Zolfaghari, L. Wen, N. Petrovic, J. Lin, and A. Knoll, “Generative ai for ocl constraint generation: Dataset collection and llm fine-tuning,” in 2024 IEEE International Symposium on Systems Engineering (ISSE), 2024, pp. 1–8

    2024

  22. [26]

    Specbook-copilot – efficient formalization of requirements using artificial intelligence in the development of mb.os,

    M. Obstbaum, I. Wior, J. Kasper, P. Vaudrevange, S. Khan, M. Keck- eisen, M. Staib, R. Freitag, and J. Schneider, “Specbook-copilot – efficient formalization of requirements using artificial intelligence in the development of mb.os,” in VDI Congress ELIV (Electronics In Vehicl...

    2024

  23. [27]

    Chat2Scenario: Scenario Extraction From Dataset Through Utilization of Large Language Model,

    Y . Zhao, W. Xiao, T. Mihalj, J. Hu, and A. Eichberger, “Chat2Scenario: Scenario Extraction From Dataset Through Utilization of Large Language Model,” in 2024 IEEE Intelligent Vehicles Symposium (IV) , Jun. 2024, pp. 559–566, arXiv:2404.16147 [cs]. [Online]. Available: http://...

    2024 arXiv

  24. [28]

    TARGET: Automated Scenario Generation from Traffic Rules for Testing Autonomous Vehicles,

    Y . Deng, J. Yao, Z. Tu, X. Zheng, M. Zhang, and T. Zhang, “TARGET: Automated Scenario Generation from Traffic Rules for Testing Autonomous Vehicles,” Oct. 2023, arXiv:2305.06018 [cs]. [Online]. Available: http://arxiv.org/abs/2305.06018

    2023 arXiv

  25. [29]

    LMM-enhanced Safety-Critical Scenario Generation for Autonomous Driving System Testing From Non-Accident Traffic Videos,

    H. Tian, X. Han, Y . Zhou, G. Wu, A. Guo, M. Cheng, S. Li, J. Wei, and T. Zhang, “LMM-enhanced Safety-Critical Scenario Generation for Autonomous Driving System Testing From Non-Accident Traffic Videos,” Jan. 2025, arXiv:2406.10857 [cs]. [Online]. Available: http://arxiv.org/a...

    2025 arXiv

  26. [30]

    LeGEND: A Top-Down Approach to Scenario Generation of Autonomous Driving Systems Assisted by Large Language Models,

    S. Tang, Z. Zhang, J. Zhou, L. Lei, Y . Zhou, and Y . Xue, “LeGEND: A Top-Down Approach to Scenario Generation of Autonomous Driving Systems Assisted by Large Language Models,” Sep. 2024, arXiv:2409.10066 [cs]. [Online]. Available: http://arxiv.org/abs/2409. 10066

    2024 arXiv

  27. [31]

    Towards specification-driven llm- based generation of embedded automotive software,

    M. S. Patil, G. Ung, and M. Nyberg, “Towards specification-driven llm- based generation of embedded automotive software,” in Bridging the Gap Between AI and Reality , B. Steffen, Ed. Cham: Springer Nature Switzerland, 2025, pp. 125–144

    2025

  28. [32]

    An empirical study of the code generation of safety-critical software using llms,

    M. Liu, J. Wang, T. Lin, Q. Ma, Z. Fang, and Y . Wu, “An empirical study of the code generation of safety-critical software using llms,” Applied Sciences , vol. 14, no. 3, 2024. [Online]. Available: https://www.mdpi.com/2076-3417/14/3/1046

    2024

  29. [33]

    On simulation-guided llm-based code generation for safe autonomous driving software,

    A. Nouri, J. Andersson, K. Hornig, Z. Fei, E. Knabe, H. Sivencrona, B. Cabrero-Daniel, and C. Berger, “On simulation-guided llm-based code generation for safe autonomous driving software,” pp. 1–11, 2025

    2025

  30. [35]

    Are requirements really all you need? a case study of llm-driven configuration code generation for automotive simulations,

    K. Lebioda, N. Petrovic, F. Pan, V . Zolfaghari, A. Schamschurko, and A. Knoll, “Are requirements really all you need? a case study of llm-driven configuration code generation for automotive simulations,” pp. 1–11, May 2025, arXiv preprint arXiv:2505.13263

    2025 arXiv

  31. [37]

    Survey of hallucination in natural language generation,

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y . Xu, E. Ishii, Y . J. Bang, A. Madotto, and P. Fung, “Survey of hallucination in natural language generation,” ACM Computing Surveys , vol. 55, no. 12, p. 1–38, Mar

  32. [38]

    Encouraging divergent thinking in large language models through multi-agent debate,

    T. Liang, Z. He, W. Jiao, X. Wang, Y . Wang, R. Wang, Y . Yang, S. Shi, and Z. Tu, “Encouraging divergent thinking in large language models through multi-agent debate,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , Y . Al-Onaizan, ...

    2024

  33. [39]

    ReConcile: Round-table conference improves reasoning via consensus among diverse LLMs,

    J. Chen, S. Saha, and M. Bansal, “ReConcile: Round-table conference improves reasoning via consensus among diverse LLMs,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , L.-W. Ku, A. Martins, and V . Srikumar,...

    2024

  34. [40]

    Rethinking the bounds of LLM reasoning: Are multi-agent discussions the key?

    Q. Wang, Z. Wang, Y . Su, H. Tong, and Y . Song, “Rethinking the bounds of LLM reasoning: Are multi-agent discussions the key?” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , L.-W. Ku, A. Martins, and V . Sr...

    2024

  35. [41]

    Self-consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. V . Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” in The Eleventh International Conference on Learning Representations , 2023. [Online]. Available: https://open...

    2023

  36. [42]

    Dynamic Self-Consistency: Leveraging Reasoning Paths for Efficient LLM Sampling,

    G. Wan, Y . Wu, J. Chen, and S. Li, “Dynamic Self-Consistency: Leveraging Reasoning Paths for Efficient LLM Sampling,” 2024, version Number: 1. [Online]. Available: https://arxiv.org/abs/2408.17017

    2024 arXiv

  37. [43]

    Internal Consistency and Self-Feedback in Large Language Models: A Survey,

    X. Liang, S. Song, Z. Zheng, H. Wang, Q. Yu, X. Li, R.-H. Li, P. Cheng, Z. Wang, F. Xiong, and Z. Li, “Internal Consistency and Self-Feedback in Large Language Models: A Survey,” Aug. 2024, arXiv:2407.14507 [cs]. [Online]. Available: http://arxiv.org/abs/2407.14507

    2024 arXiv

  38. [44]

    Escape sky-high cost: Early-stopping self-consistency for multi-step reasoning,

    Y . Li, P. Yuan, S. Feng, B. Pan, X. Wang, B. Sun, H. Wang, and K. Li, “Escape sky-high cost: Early-stopping self-consistency for multi-step reasoning,” in The Twelfth International Conference on Learning Representations , 2024. [Online]. Available: https: //openreview.net/for...

    2024

  39. [45]

    Let’s sample step by step: Adaptive-consistency for efficient reasoning and coding with LLMs,

    P. Aggarwal, A. Madaan, Y . Yang, and Mausam, “Let’s sample step by step: Adaptive-consistency for efficient reasoning and coding with LLMs,” in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , H. Bouamor, J. Pino, and K. Bali, Eds. Sing...

    2023

  40. [46]

    SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models,

    P. Manakul, A. Liusie, and M. Gales, “SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models,” in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing . Singapore: Association for Computational Lingu...

    2023

  41. [47]

    Evaluating uncertainty-based failure detection for closed-loop LLM planners,

    Z. Zheng, Q. Feng, H. Li, A. Knoll, and J. Feng, “Evaluating uncertainty-based failure detection for closed-loop LLM planners,” in ICRA 2024 Workshop on Back to the Future: Robot Learning Going Probabilistic, 2024. [Online]. Available: https://openreview.net/forum? id=9w1JnHG8Wn

    2024

  42. [48]

    CLUE: Concept- Level Uncertainty Estimation for Large Language Models,

    Y .-H. Wang, A. Bai, C.-P. Tsai, and C.-J. Hsieh, “CLUE: Concept- Level Uncertainty Estimation for Large Language Models,” Sep. 2024, arXiv:2409.03021 [cs]. [Online]. Available: http://arxiv.org/abs/2409. 03021

    2024 arXiv

  43. [49]

    Geneva, Switzerland: ISO, 2018

    International Organization for Standardization, ISO 26262: Road vehi- cles – Functional safety . Geneva, Switzerland: ISO, 2018

    2018

  44. [50]

    Cppcheck: A static analysis tool for c++,

    D. Marjam ¨aki, “Cppcheck: A static analysis tool for c++,” 2024. [Online]. Available: https://cppcheck.sourceforge.io/

    2024

  45. [51]

    Multimodal chain-of-thought reasoning in language models,

    Z. Zhang, A. Zhang, M. Li, H. Zhao, G. Karypis, and A. Smola, “Multimodal chain-of-thought reasoning in language models,” 2024

    2024

  46. [52]

    Better zero-shot reasoning with role-play prompting,

    A. Kong, S. Zhao, H. Chen, Q. Li, Y . Qin, R. Sun, X. Zhou, E. Wang, and X. Dong, “Better zero-shot reasoning with role-play prompting,”

  47. [53]

    Langprop: A code optimization framework using large language models applied to driving,

    S. Ishida, G. Corrado, G. Fedoseev, H. Yeo, L. Russell, J. Shotton, J. F. Henriques, and A. Hu, “Langprop: A code optimization framework using large language models applied to driving,” in ICLR 2024 Workshop on Large Language Model (LLM) Agents , 2024. [Online]. Available: htt...

    2024

  48. [54]

    Vecogen: Automating generation of formally verified c code with large language models,

    M. Sevenhuijsen, K. Etemadi, and M. Nyberg, “Vecogen: Automating generation of formally verified c code with large language models,”

  49. [57]

    A VL CAMEO 5™,

    A VL List GmbH, “A VL CAMEO 5™,” https://www.avl. com/en/testing-solutions/all-testing-products-and-software/ connected-development-software-tools/avl-cameo-5, 2025, accessed: 2025-04-12

    2025

  50. [58]

    (2025) Sysml.org

    SysML Partners. (2025) Sysml.org. Accessed: 2025-04-12. [Online]. Available: https://sysml.org/

    2025

  51. [60]

    Evaluation of vision language model on uml diagrams,

    A. Munde, “Evaluation of vision language model on uml diagrams,” pp. 1–11, 2025, conference presentation at Hof. [Online]. Avail- able: https://www.researchgate.net/publication/388499054 Evaluation of Vision Language Model on UML Diagrams

    2025

  52. [61]

    Beyond end-to-end vlms: Leveraging intermediate text representations for superior flowchart understanding,

    J. Ye, A. Dash, W. Yin, and G. Wang, “Beyond end-to-end vlms: Leveraging intermediate text representations for superior flowchart understanding,” December 2024, arXiv preprint arXiv:2412.16420. [Online]. Available: https://arxiv.org/abs/2412.16420

    2024 arXiv

  53. [62]

    Arrow-guided vlm: Enhancing flowchart understanding via arrow direction encoding,

    T. Omasa, R. Koshihara, and M. Morishige, “Arrow-guided vlm: Enhancing flowchart understanding via arrow direction encoding,” pp. 1–11, May 2025, arXiv preprint arXiv:2505.07864. [Online]. Available: https://arxiv.org/abs/2505.07864

    2025 arXiv

  54. [63]

    Multi-modal summarization in model-based engineering: Automotive software development case study,

    N. Petrovic, Y . Zhang, M. Maaroufi, K.-Y . Chao, L. Mazur, F. Pan, V . Zolfaghari, and A. Knoll, “Multi-modal summarization in model-based engineering: Automotive software development case study,” 2025, conference paper accepted for IntelliSys2025. [Online]. Available: https:...

    2025 arXiv

  55. [64]

    Codet5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation,

    Y . Wang, W. Wang, S. Joty, and S. C. H. Hoi, “Codet5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation,” arXiv preprint arXiv:2109.00859 , 2021

    2021 arXiv

  56. [65]

    Towards specification-driven llm-based generation of embedded automotive software,

    M. S. Patil, G. Ung, and M. Nyberg, “Towards specification-driven llm-based generation of embedded automotive software,” arXiv preprint arXiv:2411.13269, 2024

    2024 arXiv

  57. [66]

    Aegis: An advanced llm-based multi-agent for intelligent functional safety engineering,

    L. Shi, B. Qi, J. Luo, and J. Liu, “Aegis: An advanced llm-based multi-agent for intelligent functional safety engineering,” arXiv preprint arXiv:2410.12475, 2024

    2024 arXiv

  58. [67]

    Chain-of-region: Visual language models need details for diagram analysis,

    X. Li, Y . Sun, W. Cheng, Y . Zhu, and H. Chen, “Chain-of-region: Visual language models need details for diagram analysis,” pp. 1–17, January 2025, openReview preprint for ICLR 2025. [Online]. Available: https://openreview.net/pdf?id=M6fYrICcQs

    2025

  59. [68]

    Harnessing the power of large language models for automated code generation and verification,

    U. Antero, F. Blanco, J. O ˜nativia, L. M. Bergasa, R. Barea, and E. L´opez- Guede, “Harnessing the power of large language models for automated code generation and verification,” Robotics, vol. 13, no. 9, p. 137, 2024

    2024

  60. [69]

    Codebert: A pre-trained model for programming and natural languages,

    Z. Feng, D. Guo, D. Tang, N. Duan, X. Feng, M. Gong, L. Shou, B. Qin, T. Liu, and D. Jiang, “Codebert: A pre-trained model for programming and natural languages,” arXiv preprint arXiv:2002.08155 , 2020

    2002 arXiv

  61. [70]

    Startup anthropic says its new ai model can code for hours at a time,

    Anthropic, “Startup anthropic says its new ai model can code for hours at a time,” https://www.reuters.com/business/ startup-anthropic-says-its-new-ai-model-can-code-hours-time-2025-05-22/, 2025, reuters, May 22, 2025

    2025

  62. [71]

    On simulation-guided llm-based code generation for safe autonomous driv- ing software,

    A. Nouri, J. Andersson, K. D. J. Hornig, and K. Lebioda, “On simulation-guided llm-based code generation for safe autonomous driv- ing software,” arXiv preprint arXiv:2504.02141 , 2025

    2025 arXiv

  63. [72]

    Large language models are zero-shot reasoners,

    J. Kojima, S. Lu, D. Schuurmans, O. Rihaczek, L. Li, and K. Hof- mann, “Large language models are zero-shot reasoners,” arXiv preprint arXiv:2205.11916, 2022

    2022 arXiv

  64. [75]

    Chain of thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, N. Ichay, F. Xia, E. Chi, Q. V . Le, and D. Zhou, “Chain of thought prompting elicits reasoning in large language models,” in Advances in Neural Information Processing Systems (NeurIPS), 2022, available at https://arxiv.org/abs/2201.11903

    2022 arXiv

  65. [2023]

    Available: http://dx.doi.org/10.1145/3571730

    [Online]. Available: http://dx.doi.org/10.1145/3571730

    doi

  66. [2024]

    Available: https://doi.org/10.1007/s00766-023-00410-1

    [Online]. Available: https://doi.org/10.1007/s00766-023-00410-1

    doi

  67. [2025]

    Available: https://arxiv.org/abs/2411.19275

    [Online]. Available: https://arxiv.org/abs/2411.19275

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.