AOI Framework: Turning Failed Trajectories into Training Signals for Cloud Diagnosis
ORIGINAL / AOI: Turning Failed Trajectories into Training Signals for Autonomous Cloud Diagnosis
Proposes the AOI multi-agent framework that formulates autonomous operations as a trajectory learning problem under security constraints. Its novelty lies in converting failed trajectories into supervision signals via a failure trajectory closed-loop evolver, with cumulative gains validated on AIOpsLab.
01 ABSTRACT
The authors present AOI, a framework that formulates automated cloud diagnosis as a security-constrained trajectory learning problem, comprising a trainable diagnostic system, a read-write separated execution architecture, and a failure trajectory closed-loop evolver. On the AIOpsLab benchmark, the AOI runtime alone achieves 66.3% best@5; adding Observer GRPO training enables a locally deployed 14B model to reach 42.9% avg@1 on unseen fault types; the evolver converts 37 failed trajectories into diagnostic guidance, improving end-to-end avg@5 by 4.8 points and reducing variance by 35%. The authors view these components as cumulatively improving performance and enabling continual learning.
02 KEY FINDINGS
- The AOI runtime achieves 66.3% best@5 on all 86 tasks, outperforming the prior state of the art (41.9%) by 24.4 points.
- A locally deployed 14B model after Observer GRPO training reaches 42.9% avg@1 on 63 held-out tasks with unseen fault types, surpassing Claude Sonnet 4.5.
- The failure trajectory closed-loop evolver converts 37 failed trajectories into diagnostic guidance, improving end-to-end avg@5 by 4.8 points and reducing variance by 35%.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG