A research framework called PASE slashes cloud fault recovery time by more than 40% by treating repair as a code-generation problem.
Researchers introduced PASE — Planning-Aware Semantic self-healing engine — a system that uses a large language model as its core planning component. Rather than relying on static playbooks or loosely coupled AI pipelines, PASE has the LLM synthesize structured recovery plans from a library of building-block actions. A neural-symbolic world model then simulates those plans before any of them touch a live system, filtering out bad ideas before they cause more damage. A third component, a meta-prompt optimizer trained via deep reinforcement learning, tunes the prompts fed to the LLM over time, so the system keeps getting sharper. Tests on a real-world cloud fault injection dataset showed a 40%-plus reduction in average recovery time and better accuracy on fault types the system had never seen before.
Most existing fault-management pipelines bolt LLMs onto older rule-based or reinforcement learning systems without giving them a meaningful role in reasoning. PASE flips that by making the LLM the planner and using the symbolic model as a safety check — a tighter loop that other approaches don't offer. That matters because cloud outages compound fast; shaving recovery time is worth real money to anyone running infrastructure at scale.
The results come from a controlled fault injection dataset, not a live production environment, so the gap between lab numbers and on-call reality remains an open question.