Safe and Adaptive Cloud Healing Verifying LLM Generated Recovery Plans with a Neural Symbolic World Model

Jul 16, 2026

Upload by Violet

As cloud-based AI systems grow in scale and complexity, rapid fault detection and adaptive recovery become essential for reliability. The paper introduces PASE (Planning-Aware Semantic self-hEaling engine), reframing recovery as neuro-symbolic program synthesis. PASE uses an LLM as a plan synthesis engine over semantic primitives, a neural-symbolic world model to verify plan feasibility via simulation, and a DRL-trained meta-prompt optimizer to adapt prompt guidance. Experiments on real cloud fault injection show over 40% faster recovery and improved detection in unknown faults.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages13

Reading Time33 min

AccessFree to download

Introduction

Motivation and Reliability Challenges

Limits of Existing Methods

PASE Framework Overview

Safe and Adaptive Cloud Healing: Verifying LLM-Generated Recovery Plans with a Neural-Symbolic World Model arXiv :2607 .0 1595v 1 [ cs .AI] 2 Jul 2026 Junyan Tan Department of Computer Science Zhejiang University Siyuan Guo Department of Computer Science Zhejiang University Xinyue Luo Department of Computer Science Zhejiang University Haoran Lin Department of Computer Science Zhejiang University Yichen Fang Department of Computer Science Zhejiang University Tianyu Shen Department of Computer Science Zhejiang University Zeyu Qiao Department of Computer Science Zhejiang University Abstract As the scale and complexity of cloud-based AI systems continue to escalate, ensuring service reliability through rapid fault detection and adaptive recovery has become a critical challenge. While existing approaches integrate Large Language Models (LLMs) for semantic understanding and Deep Reinforcement Learning (DRL) for policy optimization, they often rely on sequential, loosely-coupled architectures that underutilize the generative and reasoning capabilities of LLMs. In this paper, we propose a paradigm shift with PASE (Planning-Aware Semantic self-hEaling engine), a novel fault self-healing framework that re-conceptualizes recovery as a neuro-symbolic program synthesis task. PASE employs an LLMas a core Plan Synthesis Engine to generate structured recovery plans from a library of semantic primitives. A Neural-Symbolic World Model verifies plan feasibility through simulation, while a Meta-Prompt Optimizer, trained via DRL, learns to generate optimal prompts that guide the LLM’s planning process. This tight “reason-plan-verify-adapt” loop enables dynamic, context-aware recovery strategy generation beyond predefined action spaces. Experiments on a real-world cloud fault injection dataset demonstrate that PASE significantly outperforms state-of-the-art methods, reducing average system recovery time by over 40% and improving fault detection accuracy in unknown fault scenarios. Our framework advances autonomous system management by unifying LLM-based reasoning with model-assisted verification and meta-learned guidance. Preprint. Figure 1: Motivation. From state→action selection without verification (late, risky recovery) to plan→ verify→ adapt: PASE synthesizes recovery plans with an LLM planner, screens them via a neural-symbolic world model, and improves planning through DRL-based meta-prompt adaptation. 1 Introduction Modern cloud-based AI systems support a wide range of mission-critical services, from online inference to large-scale distributed data processing, where even short interruptions can translate into substantial service degradation and economic loss [18, 34] . In practice, these systems are increasingly built on microservices, and their reliability hinges on heterogeneous components (e.g., container orchestrators, inference accelerators, caches, and load balancers) that are loosely coupledin implementation yet strongly dependent in execution [32, 35] . As a result, local anomalies may propagate through dependency chains and trigger cascading failures, making real-time fault management particularly challenging [25, 36] . Despite extensive efforts, accurately capturing such graph-like dependencies and their time-varying interactions remains an open problem in production-scale environments [8, 19, 26] . Fast and reliable recovery therefore becomes a core requirement for keeping these services available and for controlling operational costs [6] . However, the fault space in cloud AI systems is not only large but also heterogeneous: symptoms are distributed across logs, metrics, and alarms, and the mapping from observations to root causes is often non-bijective and context-dependent [12, 16] . For example, a memory leak can be reported as elevated CPU usage, request timeouts, or abnormal garbage collection behavior, depending on workload and resource contention [1] . Traditional rulebased ap

FAQs

What problem does PASE address in cloud-based AI systems?

PASE targets the challenge of ensuring fast, reliable, adaptive recovery when cloud services suffer faults that can propagate across loosely coupled microservice dependencies.

How does PASE generate and evaluate recovery plans?

PASE uses an LLM to synthesize structured recovery plans from semantic primitives, then checks feasibility through a neural-symbolic world model with lightweight simulation before execution.

How does DRL contribute to PASE’s recovery performance?

A DRL-based Meta-Prompt Optimizer learns prompt embeddings to steer the LLM toward higher-quality planning as the environment changes.

Jul 16, 2026

Upload by Violet