Sadat S.; Mohellebi H.; Sentouh C.; Vaillant E.; Arnoux E.; Régnier S. and Popieul J.-C. A Comparative Evaluation of LLM Prompting Strategies for OpenSCENARIO Generation from ADScene Platform In: Proceedings of the Driving Simulation Conference 2026 Europe XR, Driving Simulation Association, Antibes, France, 2026, pp. 175 - 182
Download .txt file
@inproceedings{Sadat2026,
title = {A Comparative Evaluation of LLM Prompting Strategies for OpenSCENARIO Generation from ADScene Platform},
author = {Sofiane Sadat and Hakim Mohellebi and Chouki Sentouh and Eric Vaillant and Emmanuel Arnoux and Stéphane Régnier and Jean-Christophe Popieul},
editor = {Andras Kemeny and Jean-Rémy Chardonnet and Florent Colombet and Stéphane Espié},
doi = {https://doi.org/10.82157/dsa/2026/25},
isbn = {978-2-9573777-9-4},
year = {2026},
date = {2026-09-16},
booktitle = {Proceedings of the Driving Simulation Conference 2026 Europe XR},
volume = {11},
pages = {175 - 182},
address = {Antibes, France},
organization = {Driving Simulation Association},
abstract = {The rapid progress of Generative Artificial Intelligence (GenAI) has created new opportunities for automating engineering workflows (Gozalo-Brizuela and Merchan, 2024), including simulation scenario generation for the verification and validation of Autonomous Driving Systems (ADS). In this context, this study investigates the automatic generation of Python OpenSCENARIO code (ASAM, 2026) from structured ADScene (Guyonvarch et al., 2023) descriptions using Large Language Models (LLMs). Building on the limitations identified in earlier multi-agent approaches (Mohellebi, 2025), this work explores several alternative generation strategies, including direct single-step generation, retrieval-augmented generation, structured constrained generation, reasoning-based class selection, and full-documentation prompting. To assess these methods, we introduce a hierarchical evaluation framework based on executability, behavioral fidelity, and trigger fidelity, complemented by the number of retries required to obtain a correct generation. The results show that simpler generation pipelines consistently improve executability compared with the multi-agent baseline, while documentation grounding plays a central role in improving behavioral correctness and trigger conformity. Among the evaluated methods, the fulldocumentation variant provides the strongest overall results while remaining straightforward to implement.
Overall, the findings suggest that, for this task, simpler generation strategies are more effective than complex multi-agent pipelines.
},
keywords = {},
}
Download .bib file
TY - CONF
TI - A Comparative Evaluation of LLM Prompting Strategies for OpenSCENARIO Generation from ADScene Platform
AU - Sadat, Sofiane
AU - Mohellebi, Hakim
AU - Sentouh, Chouki
AU - Vaillant, Eric
AU - Arnoux, Emmanuel
AU - Régnier, Stéphane
AU - Popieul, Jean-Christophe
C1 - Antibes, France
C3 - Proceedings of the Driving Simulation Conference 2026 Europe XR
DA - 2026/09/16
PY - 2026
SP - 175
EP - 182
LA - en-US
PB - Driving Simulation Association
SN - 978-2-9573777-9-4
L2 - https://proceedings.driving-simulation.org/proceeding/dsc-2026/a-comparative-evaluation-of-llm-promptingstrategies-for-openscenario-generation-fromadscene-platform
ER -
Download .ris file