一句话：SAFE 用 RAG、知识引导提示、CoT 与自校验从多模态碰撞数据重建可复现场景，提升 ADS 测试的违规与真实事故检出。
从多模态碰撞报告中提取真实且可复现的测试场景，是保障自动驾驶系统（ADS）安全的关键难题；而大语言模型（LLM）在直接生成场景时易产生幻觉并丢失地图结构，导致道路布局与车辆行为不真实。为此，作者提出 SAFE——一个基于场景的 ADS 测试框架，通过多模态提取，结合检索增强生成（RAG）、知识引导提示、思维链（CoT）推理与自校验，改进从多模态碰撞数据中的场景重建。评估显示，SAFE 在路网细节提取上达到 93.8% 准确率，参与者信息 80.0%，环境上下文 100%；在人类研究中，其重建的路网与车辆行为一致性优于 LCTGen 与 AC3R。在相同 ADS 与仿真器设置下，SAFE 比 LCTGen 与 AC3R 分别多检出 39 与 71 个安全违规，比 LCTGen 多复现 12 个真实碰撞案例；在 AC3R 支持的 19 个案例上额外复现 1 个，五次运行具统计显著提升。SAFE 在 25 秒内生成场景，在 MetaDrive 中仅需 1 个案例（IDM）和 3 个案例（PPO）、在 BeamNG 中仅需 1 个案例（Auto）即可触发违规。
Ensuring the safety of Autonomous Driving Systems (ADS) requires realistic and reproducible test scenarios, yet extracting such scenarios from multimodal crash reports remains a major challenge. Large Language Models (LLMs) often hallucinate and lose map structure, resulting in unrealistic road layouts and vehicle behaviors. To address this, we introduce SAFE, a novel Scenario-based ADS testing Framework via multimodal Extraction, which leverages Retrieval-Augmented Generation (RAG), knowledge-grounded prompting, Chain-of-Thought (CoT) reasoning, and self-validation to improve scenario reconstruction from multimodal crash data. SAFE achieves 93.8% accuracy in extracting road network details, 80.0% for actor information, and 100% for environmental context. In human studies, SAFE outperforms LCTGen and AC3R in reconstructing consistent road networks and vehicle behaviors. Under identical ADS and simulator settings, SAFE detects 39 and 71 more safety violations than LCTGen and AC3R, respectively, and reproduces 12 more real-world crash cases than LCTGen. On 19 cases supported by AC3R, SAFE reproduces one additional crash case with statistically significant gains across five runs. It generates scenarios within 25 seconds and triggers violations after just 1 case (IDM) and 3 cases (PPO) in MetaDrive, as well as 1 case (Auto) in BeamNG.
@inproceedings{paperbot11650,
title = {SAFE: Harnessing LLM for Scenario-Driven ADS Testing from Multimodal Crash Data},
author = {Siwei Luo and Yang Zhang and Yao Deng and Linfeng Liang and Xi Zheng},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3744916.3773232}
}