Technical Summary: Safe and Adaptive Cloud Healing with PASE

1. Problem Statement

Modern cloud-based AI systems, characterized by microservices and heterogeneous components (e.g., container orchestrators, inference accelerators), face significant challenges in maintaining reliability. Local anomalies often propagate through dependency chains, triggering cascading failures. The core difficulties in fault management include:

• Heterogeneity and Complexity: Fault symptoms are distributed across logs, metrics, and alarms, with non-bijective mappings between observations and root causes.
• Limitations of Existing Approaches:
  • Rule-based systems rely on static thresholds and fail to generalize to evolving deployments or unseen fault patterns.
  • Model-based techniques (e.g., Bayesian networks) become prohibitively expensive to maintain as systems scale and change.
  • Current LLM-DRL Hybrids: While combining Large Language Models (LLMs) for semantic understanding and Deep Reinforcement Learning (DRL) for policy optimization shows promise, existing implementations are often sequential and loosely coupled. They typically treat the LLM as a passive state interpreter and the DRL agent as a downstream policy selector operating within a predefined action space. This design underutilizes the LLM's generative planning capabilities and lacks a dedicated pre-execution verification stage, leading to brittle recovery actions, especially under novel fault conditions.

2. Methodology: The PASE Framework

The paper proposes PASE (Planning-Aware Semantic self-hEaling engine), a paradigm shift that reconceptualizes cloud fault recovery as a neuro-symbolic program synthesis task. Instead of selecting single-step actions, PASE generates multi-step recovery plans through a tight "reason-plan-verify-adapt" loop.

2.1 Unified Semantic Observation and Plan Synthesis

• Semantic Scene Description: Raw heterogeneous observations ($O_t$), comprising logs, metrics, and alarms, are mapped by an LLM into a structured natural language description ($D_t$). This preserves relational and causal information often lost in vectorization.
• Plan Synthesis Engine (PSE): An LLM acts as the core planner. Given the semantic description $D_t$ and a library of parameterizable Recovery Primitives ($R$) (e.g., Restart, ScaleOut, Reroute), the PSE generates a structured recovery plan $\Pi_t$. This plan is a sequence or graph of primitives with instantiated parameters, allowing for dynamic, multi-step strategies tailored to specific fault contexts.

2.2 Neural-Symbolic World Model (NSWM) for Verification

To mitigate the risks of hallucination and inconsistency in LLM-generated plans, PASE employs a Neural-Symbolic World Model.

• Function: The NSWM learns the stochastic dynamics of the cloud system at the level of recovery primitives. It predicts state transitions and expected changes in health indicators ($\Delta H$) when a primitive is applied.
• Simulation & Scoring: Before execution, the NSWM performs a rollout simulation of the candidate plan $\Pi_t$. It calculates a Plan Feasibility Score ($F(\Pi_t)$), defined as a negative weighted sum of predicted resource violation risks and expected residual fault severity.
• Decision: Plans are only executed if $F(\Pi_t)$ exceeds a safety threshold $\tau$, ensuring reliability prior to real-world intervention.

2.3 Meta-Prompt Optimization via Reinforcement Learning

To adapt to new fault patterns without costly full-model fine-tuning, PASE introduces a Meta-Prompt Optimizer (MPO).

• Mechanism: A DRL agent (using Soft Actor-Critic, SAC) learns to generate continuous prompt embeddings ($p_t$) that condition the PSE.
• Reward Signal: The MPO is optimized based on a composite reward including recovery efficiency, resource efficiency, and a shaped reward if the NSWM's pre-execution score was high.
• Adaptation: This allows the system to rapidly adjust the LLM's reasoning trajectory by learning optimal prompting strategies for unseen fault types, rather than relearning recovery policies from scratch.

2.4 Training and Execution Loop

The framework operates through three coordinated phases:

1. NSWM Pre-training: Trained on historical logs to predict primitive outcomes.
2. MPO Warm-up: The MPO is trained in simulation with a fixed base prompt.
3. Joint Fine-tuning: The entire loop is fine-tuned with real-world interactions.
   During online operation, the system generates a description, produces a prompt, synthesizes a plan, verifies it via the NSWM, and executes it only if feasible.

3. Key Contributions

1. Neuro-Symbolic Formulation: The paper formulates cloud fault recovery as a program synthesis problem, shifting the decision unit from single-step action selection to multi-step plan generation.
2. Verified Planning Loop: It designs a closed loop coupling an LLM-based synthesizer with a neural-symbolic world model for feasibility screening, improving reliability before execution.
3. Meta-Learned Guidance: It introduces a DRL-based meta-prompt optimizer that adapts prompting in a continuous embedding space, enabling efficient online adaptation to unseen fault types.

4. Experimental Results

Experiments were conducted on a real-world OpenStack-based cloud fault injection dataset ("Failure-Dataset-OpenStack"), comparing PASE against baselines including IFSHM (LLM+DRL), Deformable DETR-FD, GCN-FR, and Rule-Based systems.

• Recovery Efficiency: PASE reduced the Average System Recovery Time (MTTR) to 72.3 seconds, a 40% improvement over the IFSHM baseline (85.1s) and significantly better than other methods.
• Fault Detection: PASE achieved a Fault Detection F1-score of 0.942, outperforming the IFSHM baseline (0.918) and other state-of-the-art methods.
• Safety and Verification: The NSWM's feasibility score showed a 92% correlation with actual plan success rates. The "Direct Execution (No Verify)" variant resulted in a significantly lower safety score (0.58) compared to the full PASE (0.89).
• Adaptability: When faced with a novel "hybrid CPU-Memory deadlock fault," PASE (via MPO) increased recovery success from 40% to over 80% within 15 interaction episodes. In contrast, the IFSHM baseline required full prompt-tuning and converged at a lower performance (75%).
• Ablation Studies: Removing the NSWM increased MTTR to 97.8s and lowered the safety score to 0.65. Removing the Meta-Prompt Optimizer increased adaptation steps from 15 to 38, confirming the necessity of both verification and meta-optimization.

5. Significance and Claims

The paper claims that PASE advances autonomous system management by unifying LLM-based reasoning with model-assisted verification and meta-learned guidance. Its primary significance lies in:

• Paradigm Shift: Moving beyond using LLMs as passive encoders or DRL as mere policy selectors. PASE positions the LLM as an active planning engine that generates structured recovery programs.
• Safety and Reliability: By integrating a neural-symbolic world model, the framework addresses the "hallucination" risks of LLMs, ensuring that recovery actions are verified for feasibility before execution.
• Rapid Adaptation: The meta-prompt optimization strategy allows the system to adapt to previously unseen fault types with significantly fewer interaction episodes compared to traditional fine-tuning or static prompting methods.

The authors acknowledge limitations, noting that the current LLM planner relies on offline fine-tuning and that the meta-prompt optimizer requires considerable interaction data. Future work is proposed to focus on online continual learning, model compression for edge-cloud deployment, and extending the framework to federated multi-cloud environments.