一句话：提出 InsightTriage，用 LLM 联合分析用户查询与车辆日志，实现可解释的自动化事件分级
智能网联汽车车载系统日益复杂，实际使用中异常频发，而人工事件分级难以应对规模与复杂度增长。本文提出 InsightTriage，一个面向真实工业场景设计并部署的端到端自动事件分级系统，利用大语言模型联合分析用户查询与车辆日志，生成准确且可解释的分级结果。针对领域知识缺失，系统在离线阶段借助 LLM 从历史车辆日志中自动抽取知识，构建组件级结构化知识库；在线阶段为缓解 LLM 的上下文长度限制，引入以对比学习预训练的日志检索器，高效定位与查询相关的日志条目。在来自某全球顶级智能网联汽车供应商生产环境的数据集上，InsightTriage 的加权 F1 达 0.801，比最佳基线提升 0.137，并能输出可解释的分级报告。
With the rapid development of intelligent and connected vehicles (ICVs), in-vehicle systems have grown increasingly complex, leading to frequent anomalies in real-world usage. To maintain a high-quality user experience, efficient incident triage is essential for timely issue resolution. However, with the growing volume and complexity of incidents, traditional manual triage processes are becoming increasingly insufficient. Recent advances in large language models (LLMs) have demonstrated strong capabilities in reasoning and natural language understanding, offering new opportunities for automation. We propose InsightTriage, an end-to-end automated incident triage system designed and deployed for real-world industrial scenarios, which leverages LLMs to jointly analyze user queries and vehicle logs to generate accurate and explainable triage results. To address the lack of domain-specific knowledge, InsightTriage first constructs a component-level structured knowledge base in an offline phase by leveraging LLMs to automatically extract knowledge from historical vehicle logs. In the online phase, to mitigate the context length limitations of LLMs in log analysis, InsightTriage introduces a log retriever pretrained with contrastive learning, which efficiently identifies query-relevant log entries. Comprehensive experiments conducted on a dataset collected from the production environment of a top-tier global supplier for ICVs, demonstrate that InsightTriage achieves superior performance, reaching a weighted F1-Score of 0.801 (an improvement of 0.137 over the best baseline), while also providing explainable triage reports.
@inproceedings{paperbot12177,
title = {LLM-Assisted Joint Ticket and Log Analysis for Incident Triage in Intelligent and Connected Vehicles},
author = {Ruowei Fu and Shenglin Zhang and Wenwei Gu and Weiguo Li and Yongqian Sun and Dan Pei},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834521}
}