华为 GTS 提出 NetCanvas：为排障 Agent 引入可交互视觉拓扑，CTBench 通过率升至 54.5%
华为 GTS 用一张「活的」拓扑图帮排障 Agent 记住网络结构，复杂场景准确率从约 10% 提到约 90%，说明 Agent 的瓶颈常在外围系统而非模型本身。
华为 GTS 团队发现，纯文本 Agent 在真实网络排障中会「拓扑失忆」——记得住单条命令，却记不住整张网长什么样，于同一节点反复打转。他们提出 NetCanvas，用可交互的视觉拓扑图作为 Agent 的外部工作记忆，边查边画。CTBench 上综合通过率从 30.3% 升至 54.5%，双防火墙等复杂任务从约 10% 升至约 90%。
正文摘录
- title: Huawei GTS Teaches Agents to "Diagnose Networks by Looking at Them," Nearly Cracking the Dual-Firewall Challenge - sourcecompany: QbitAI - bodymarkdown: "Intelligence is not confined to the brain alone. When we use pen and paper or maps to untangle complex problems, these external tools become our 'extended mind.'" True intelligence knows how to use external diagrams for "cognitive offloading." The IP bearer network is the foundation of digital systems across every industry. A single card swipe payment, a single online appointment registration—none of it works without it. Yet once this network fails, the clues are often buried deep in massive volumes of alerts or silent configurations. When business is disrupted, the first hurdle in troubleshooting is often not how to fix it, but …