PASE: LLM-driven Cloud Healing that Verifies Recovery Plans

Junyan Tan and six co-authors submitted an arXiv paper on 2 Jul 2026 proposing PASE, a Planning-Aware Semantic self-healing engine that treats cloud recovery as a neuro-symbolic program synthesis task. Experiments on a real-world cloud fault injection dataset show PASE reduces average system recovery time by over 40% and improves fault detection accuracy in unknown fault scenarios.

What is PASE?

PASE is a self-healing framework that reconceptualizes recovery as program synthesis, using an LLM as the Plan Synthesis Engine, a Neural-Symbolic World Model to verify plans by simulation, and a Meta-Prompt Optimizer trained via Deep Reinforcement Learning to guide the LLM. The approach frames recovery plans as structured outputs built from a library of semantic primitives rather than as fixed action sequences.

PASE aims to move beyond sequential, loosely coupled architectures that mix semantic understanding and policy optimization. The paper positions the system as a tight reason-plan-verify-adapt loop that produces dynamic, context-aware recovery strategies outside predefined action spaces.

How does PASE verify and adapt LLM-generated plans?

PASE first uses an LLM Plan Synthesis Engine to generate structured recovery plans from a library of semantic primitives, then runs those plans through a Neural-Symbolic World Model that verifies feasibility via simulation, while a Meta-Prompt Optimizer refines prompts to the LLM. The loop is explicit: plan synthesis, simulated verification, and meta-prompt-driven adaptation.

The Neural-Symbolic World Model provides model-assisted verification rather than relying solely on the LLM’s internal reasoning. The Meta-Prompt Optimizer is trained with DRL to produce prompts that steer the LLM toward better plans, closing the reason-plan-verify-adapt loop described by the authors.

How did PASE perform in experiments?

On a real-world cloud fault injection dataset used by the authors, PASE "significantly outperforms state-of-the-art methods," cutting average system recovery time by over 40% and improving fault detection accuracy in unknown fault scenarios. The paper reports these results as the empirical evidence supporting the proposed neuro-symbolic program synthesis paradigm.

The submission is 13 pages long and presents the experimental comparison against prior approaches that combine LLMs and Deep Reinforcement Learning, arguing that existing methods underutilize LLM generative and reasoning capabilities due to looser coupling between components.

Why it matters

PASE shifts the design of autonomous cloud recovery from fixed action policies toward LLM-guided program synthesis with model-based verification. That matters because cloud incidents often present unexpected fault types; the authors show PASE improves handling of unknown faults and cuts recovery time by a measurable margin. For operations teams, a system that verifies LLM plans through simulation reduces reliance on ad hoc manual checks and tightens the safety envelope around automated actions.

What to watch

Look for the paper’s code and datasets linked from the arXiv entry and for any follow-up work that opens the Meta-Prompt Optimizer or the Neural-Symbolic World Model implementations. The next concrete signal will be replication of the reported "over 40%" recovery-time reduction on other cloud fault injection datasets or in live production trials.

Notes

The paper is titled "Safe and Adaptive Cloud Healing: Verifying LLM-Generated Recovery Plans with a Neural-Symbolic World Model," submitted 2 Jul 2026 to arXiv (arXiv:2607.01595) by Junyan Tan, Haoran Lin, Siyuan Guo, Yichen Fang, Xinyue Luo, Tianyu Shen and Zeyu Qiao.