Primary sourcearXiv cs.AI
View sourceEvent ·Multi-source
MCP-Injector研究:工具故障下120B模型容错并不优于20B
arXiv 论文 MCP-Injector 是一款针对 MCP 工具调用路径的协议级故障注入中间件,支持永久、瞬时、间歇三类故障模式。故障模型基于 1,774 个开源 MCP 仓库中 57,240 条 issue 与 PR 记录标定。在 1,350 组实验中,两个 gpt-oss 骨干(20B 与 120B)在故障注入下任务完成分从 8.08 跌至 3.23,120B 的容错率与 20B 无显著差异(p=0.96)。结果显示 OpenAI Agents SDK、LangChain 等框架层的错误处理对容错影响大于模型规模。
Current read
给做智能体的朋友:这篇测了 MCP 工具挂掉时各家 agent 框架的表现,结论挺反直觉,模型变大不顶用,框架的错误处理才关键。
11 sources34° AI heatLast updated 2026/10/23 12:35:51
Evidence
11 sourcesRelated source 1TestingCatalog
View sourceRelated source 2kimmonismus
View sourceRelated source 3rohanpaul_ai
View sourceRelated source 4IT之家
View sourceRelated source 5VnExpress Số hóa
View sourceRelated source 6Artificial Analysis
View sourceRelated source 7Teortaxes
View sourceRelated source 8theverge
View sourceRelated source 9@koltregaskes
View sourceRelated source 10Decoder
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.