arXiv：cs.LG· Rudrendu Kumar Paul, Sourav Nandy·· 5 小时前AI 评分56
复合AI系统可靠性研究：从150起生产事故归纳23种失效模式与韧性模式目录
Compound AI System Reliability: A Failure Taxonomy and Resilience Pattern Catalog from 150 Production Incidents
AI 导读
研究者分析150份来自开源复合AI项目和匿名企业部署的生产事故报告，构建出23种失效模式的分类体系，分为检索、生成、工具、编排和集成五类。故障注入实验显示，断路器可减少89%的级联传播，输出质量门可拦截73%的静默质量退化，组件隔离可将影响范围缩小64%；实现三种及以上韧性模式的系统MTTR较无结构监控基线降低71%。
正文
Abstract:Deploying compound AI systems reliably and safely requires understanding failure modes that emerge at component boundaries, not within individual models. Cascading errors propagate across component boundaries, silent quality degradation evades standard monitoring, and coordination failures yield incorrect collective behavior from individually correct parts. We analyze 150 production incident reports from open-source compound AI projects and anonymized enterprise deployments to construct a taxonomy of 23 failure modes organized into five categories: retrieval failures, generation failures, tool failures, orchestration failures, and integration failures. For each category, we propose resilience patterns with measured effectiveness from controlled fault injection experiments. Circuit breakers reduce cascade propagation by 89%, output quality gates catch 73% of silent degradation before user impact, and component isolation reduces blast radius by 64%. Systems implementing three or more resilience patterns from our catalog reduce mean-time-to-recovery (MTTR) by 71% compared to unstructured monitoring baselines. We release the incident taxonomy and pattern catalog as a practitioner resource.
| Comments: | Accepted at the AIWILD Workshop, ICML 2026. Camera-ready version |
| Subjects: | Software Engineering (cs.SE); Artificial Intelligence (cs.AI); Distributed, Parallel, and Cluster Computing (cs.DC); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.02503 [cs.SE] |
| (or arXiv:2610.02503v1 [cs.SE] for this version) | |
| https://doi.org/10.48550/arXiv.2610.02503 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Rudrendu Kumar Paul [view email]
[v1]
Thu, 1 Oct 2026 21:25:25 UTC (25 KB)
来源：arXiv：cs.LG · arxiv.org