PASE Enhances Cloud Healing with LLM-Generated Recovery Plans

Ensuring the reliability of large-scale cloud-based AI systems is a significant challenge, particularly in detecting faults and implementing adaptive recovery. Traditional methods often combine Large Language Models (LLMs) for understanding and Deep Reinforcement Learning (DRL) for policy optimization, but these approaches can be limited by their sequential architectures. A novel framework, PASE, proposes a new paradigm by treating fault self-healing as a neuro-symbolic program synthesis task. PASE utilizes an LLM as a central Plan Synthesis Engine to create structured recovery plans from a library of semantic primitives. These plans are then verified for feasibility by a Neural-Symbolic World Model through simulation. Furthermore, a Meta-Prompt Optimizer, trained with DRL, learns to generate optimal prompts that guide the LLM's planning process, enabling dynamic and context-aware recovery strategies beyond predefined actions. Experiments show PASE significantly reduces recovery time and improves fault detection accuracy in unknown scenarios.